TensorFlow 对象检测 API 变量在检查点中不可用答案

【问题标题】：Tensorflow Object detection API Variable is not available in checkpointTensorFlow 对象检测 API 变量在检查点中不可用
【发布时间】：2020-05-31 06:13:00
【问题描述】：

我正在尝试从 Tensorflow's model zoo 重新训练对象检测模型，但我得到了一个在检查点中找不到的变量列表（请参阅下面列表中的一些警告）。

我正在使用：

Tensoflow 版本：1.15.2
Windows 10
Python 3.6.10

我使用的配置文件是带有预训练模型的 zip 文件中的配置文件，我唯一更改的是检测到的类的数量和路径。

我尝试了多个模型，但每次都遇到相同的问题。

迁移学习有效，但我想我没有使用大量的预训练权重。 是什么原因造成的？

Full log can be found here.

My config file can be found here

警告：

W0215 22:04:03.197386  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/BatchNorm/beta/ExponentialMovingAverage] is not available in checkpoint
W0215 22:04:03.197386  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/BatchNorm/beta/RMSProp] is not available in checkpoint
W0215 22:04:03.197386  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/BatchNorm/beta/RMSProp_1] is not available in checkpoint
W0215 22:04:03.197386  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/BatchNorm/gamma/ExponentialMovingAverage] is not available in checkpoint
W0215 22:04:03.197386  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/BatchNorm/gamma/RMSProp] is not available in checkpoint
W0215 22:04:03.197386  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/BatchNorm/gamma/RMSProp_1] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/weights/ExponentialMovingAverage] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/weights/RMSProp] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_0/weights/RMSProp_1] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/BatchNorm/beta/ExponentialMovingAverage] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/BatchNorm/beta/RMSProp] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/BatchNorm/beta/RMSProp_1] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/BatchNorm/gamma/ExponentialMovingAverage] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/BatchNorm/gamma/RMSProp] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/BatchNorm/gamma/RMSProp_1] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/depthwise_weights/ExponentialMovingAverage] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/depthwise_weights/RMSProp] is not available in checkpoint
W0215 22:04:03.198379  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_depthwise/depthwise_weights/RMSProp_1] is not available in checkpoint
W0215 22:04:03.199381  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_pointwise/BatchNorm/beta/ExponentialMovingAverage] is not available in checkpoint
W0215 22:04:03.199381  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_pointwise/BatchNorm/beta/RMSProp] is not available in checkpoint
W0215 22:04:03.199381  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_pointwise/BatchNorm/beta/RMSProp_1] is not available in checkpoint
W0215 22:04:03.199381  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_pointwise/BatchNorm/gamma/ExponentialMovingAverage] is not available in checkpoint
W0215 22:04:03.199381  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_pointwise/BatchNorm/gamma/RMSProp] is not available in checkpoint
W0215 22:04:03.199381  5668 variables_helper.py:157] Variable [FeatureExtractor/MobilenetV1/Conv2d_10_pointwise/BatchNorm/gamma/RMSProp_1] is not available in checkpoint

【问题讨论】：

标签： tensorflow object-detection transfer-learning

【解决方案1】：

不确定这是否仍然相关，但我也遇到了这个问题，看起来这可能是模型存储库的贡献者之一here 所述的显示错误。

无论基础模型是什么，都会出现这个bug，因为我们使用预训练模型来训练ssd。如果我们随机初始化ssd模型，这不会发生。而且，它会比随机初始化时使用更多的cpu训练。（大约 55% 的 CPU 使用率相比 45%）

【讨论】：