docker中tensorflow object_detection API训练问题及解决方法

1.尝试运行models/research/object_detection/目录下的.py文件时出现：

ModuleNotFoundError: No module named 'object_detection'

需在models/research/目录下执行：

python setup.py install

出现ModuleNotFoundError: No module named 'pycocotools'

执行：

conda install pycocotools

出现 ImportError: No module named 'nets'

需在models/research/slim目录下执行：

python setup.py install

2.训练中出现的问题

① raise ValueError('First step cannot be zero.')ValueError: First step cannot be zero.

　　删除预训练模型Config中的下列代码

1 2 3 4

schedule{         step:0,         learning_rate:0.00001 }

 ②Assertion failed: [maximum box coordinate value is larger than 1.100000: ] [1.1435889]

问题：标注数据，超出了图像的长宽

解决方法：在生成标注数据时，将x，y坐标进行截断

3.出现 ValueError: Tried to convert 't' to a tensor and failed. Error: Argument must be a dense tensor: range(0, 3) - got shape [3], but wanted [].错误

找到错误提示中的learning_schedules.py

将其167～19行

    rate_index = tf.reduce_max(tf.where(tf.greater_equal(global_step, boundaries),
                                          range(num_boundaries),
                                          [0] * num_boundaries))

修改成：

      rate_index = tf.reduce_max(tf.where(tf.greater_equal(global_step, boundaries),
                                          list(range(num_boundaries)),
                                          [0] * num_boundaries)