我正在为我们的项目使用“ Tensorflow对象检测API”。
经过多次尝试和错误，我成功发明了发明， 但是当我测试模型时，我认为我可以改善模型的分析时间
我的电脑是
“ Ubuntu 16.04”，
“ i7-7700k”，
“单个GPU，GTX1080ti”，
“ CUDA 9.1”，“ cuDNN 7.0.5”，
模型的平均分析时间为每张img 1.1秒（尺寸1920 * 1080，jpg格式）
由于给出了Tensorflow对象检测API教程的测试代码，因此它似乎是“读取图像文件并将其转换为np数组，而不是tf记录”
那么，...如果将test_imgs作为tf记录格式输入，可以改善模型的分析时间吗？
如果是，该怎么办？
请帮助我或给出一些提示。
这是我的代码
TEST_IMAGE_PATHS = [ os.path.join(PATH_TO_TEST_IMAGES_DIR, 'image{}.jpg'.format(i)) for i in range(start_idx, end_idx) ]
with self.detection_graph.as_default():
with tf.Session(graph=self.detection_graph) as sess:
# Definite input and output Tensors for detection_graph
image_tensor = self.detection_graph.get_tensor_by_name('image_tensor:0')
# Each box represents a part of the image where a particular object was detected.
detection_boxes = self.detection_graph.get_tensor_by_name('detection_boxes:0')
# Each score represent how level of confidence for each of the objects.
# Score is shown on the result image, together with the class label.
detection_scores = self.detection_graph.get_tensor_by_name('detection_scores:0')
detection_classes = self.detection_graph.get_tensor_by_name('detection_classes:0')
num_detections = self.detection_graph.get_tensor_by_name('num_detections:0')
total_size = len(TEST_IMAGE_PATHS)
cur_idx = 0
while True:
if total_size <= 0:
break
cur_lst = []
for i in range(min(my_batch_size, total_size)):
image_path = TEST_IMAGE_PATHS[cur_idx + i]
try:
image = Image.open(image_path)
except FileNotFoundError:
print('[{}] does not exist'.format(image_path))
continue
image_np = self.load_image_into_numpy_array(image)
cur_lst.append(image_np)
cur_feed_lst = np.asarray(cur_lst)
(boxes, scores, classes, _) = sess.run([detection_boxes, detection_scores, detection_classes, num_detections],\
feed_dict={image_tensor: cur_feed_lst})
myboxes = np.squeeze(boxes)
myclasses = np.squeeze(classes).astype(np.int32)
myscores = np.squeeze(scores)
im_width, im_height = (1920, 1080)
if cur_feed_lst.shape[0] == 1:
tmp_bbox_list = []
for i in range(min(20, myboxes.shape[0])):
if myscores[i] >= self.detect_threshold:
ymin, xmin, ymax, xmax = tuple(myboxes[i].tolist())
(left, right, top, bottom) = (int(xmin * im_width), int(xmax * im_width),
int(ymin * im_height), int(ymax * im_height))
# score, classN, left, right, top, bottom
curclassn = self.category_index[myclasses[i]]['name']
cur_bbox_data = '{0:.3f}\t{1}\t{2}\t{3}\t{4}\t{5}'.format(myscores[i], curclassn, left, top, right, bottom)
tmp_bbox_list.append(cur_bbox_data)
if not tmp_bbox_list:
pass
else:
image_path = TEST_IMAGE_PATHS[cur_idx]
res_list[image_path] = tmp_bbox_list
vis_util.visualize_boxes_and_labels_on_image_array(
cur_feed_lst[0],
np.squeeze(boxes),
np.squeeze(classes).astype(np.int32),
np.squeeze(scores),
self.category_index,
use_normalized_coordinates=True,
line_thickness=8)
pass
else:
for j in range(min(my_batch_size, total_size)):
myboxes2 = myboxes[j]
myclasses2 = myclasses[j]
myscores2 = myscores[j]
tmp_bbox_list = []
for i in range(min(20, myboxes2.shape[0])):
if myscores2[i] >= self.detect_threshold:
ymin, xmin, ymax, xmax = tuple(myboxes2[i].tolist())
(left, right, top, bottom) = (int(xmin * im_width), int(xmax * im_width),
int(ymin * im_height), int(ymax * im_height))
# score, classN, left, right, top, bottom
curclassn = self.category_index[myclasses2[i]]['name']
cur_bbox_data = '{0:.3f}\t{1}\t{2}\t{3}\t{4}\t{5}'.format(myscores2[i], curclassn, left, top, right, bottom)
tmp_bbox_list.append(cur_bbox_data)
if not tmp_bbox_list:
pass
else:
image_path = TEST_IMAGE_PATHS[cur_idx + j]
res_list[image_path] = tmp_bbox_list
pass
# update
cur_idx += my_batch_size
total_size -= my_batch_size
def load_image_into_numpy_array(self, image):
(im_width, im_height) = image.size
return np.array(image.getdata()).reshape((im_height, im_width, 3)).astype(np.uint8)