[实习2021-03-30]Object_detection_tutorial踩坑记录

Object_detection_tutorial踩坑记录

环境记录

• Ubuntu
• tensorflow-gpu == 2.4.1
• python == 3.6.13

issue 1

AttributeError: 'AutoTrackable' object has no attribute 'output_shapes'

解决方案

# masking_model.output_shapes 替换为以下
masking_model.signatures['serving_default'].output_shapes

issue 2

在运行maskrcnn模型时出现
TypeError: Cannot interpret 'tf.uint8' as a data type

#~/miniconda3/envs/rcnn/lib/python3.6/site-packages/object_detection/utils/ops.py in reframe_box_masks_to_image_masks(box_masks, boxes, image_height, image_width, resize_method)
resize_method = 'nearest' if box_masks.dtype == tf.uint8 else resize_method # 这一行报错

问题原因

numpy.ndarray无法与tensor比较

解决方案

1. 在进入reframe_box_masks_to_image_masks函数之前将box_masks转换为Tensor格式

   tf.convert_to_tensor(output_dict[‘detection_masks’])s

2. 修改自己环境下的site-packages/object_detection/utils/ops.py的源码

   1. import numpy as np

   2. resize_method = ‘nearest’ if box_masks.dtype == tf.uint8 else resize_method

      修改为

      resize_method = ‘nearest’ if box_masks.dtype == np.uint8 else resize_method