我的模型faster_rcnn_inception_v2_coco在TensorFlow对象检测API中找到，可以从官方仓库here下载。
当我使用tensorflow-gpu 1.13.1并简单地通过Docker安装w /找到的说明here时，一切都可以完美地进行推断。
但是当我尝试运行从源构建的tensorflow-gpu（运行TF 1.13.1）（通过Docker并附有指令here）时， CPU优化的优势尝试进行推断时，遇到以下错误。
2019-04-11 20:27:59.196073: W external/org_tensorflow/tensorflow/core/framework/op_kernel.cc:1408] OP_REQUIRES failed at conv_ops_fused_impl.h:648 : Unimplemented: Fusion is not implemented: [BiasAdd,Relu6]
2019-04-11 20:27:59.196141: E external/org_tensorflow/tensorflow/core/common_runtime/executor.cc:637] Executor failed to create kernel. Unimplemented: Fusion is not implemented: [BiasAdd,Relu6]
[[{{node Conv/Relu6}}]]
2019-04-11 20:28:01.013470: W external/org_tensorflow/tensorflow/core/framework/op_kernel.cc:1408] OP_REQUIRES failed at conv_ops_fused_impl.h:648 : Unimplemented: Fusion is not implemented: [BiasAdd,Relu6]
2019-04-11 20:28:01.013541: E external/org_tensorflow/tensorflow/core/common_runtime/executor.cc:637] Executor failed to create kernel. Unimplemented: Fusion is not implemented: [BiasAdd,Relu6]
[[{{node Conv/Relu6}}]]
2019-04-11 20:28:02.892400: W external/org_tensorflow/tensorflow/core/framework/op_kernel.cc:1408] OP_REQUIRES failed at conv_ops_fused_impl.h:648 : Unimplemented: Fusion is not implemented: [BiasAdd,Relu6]
2019-04-11 20:28:02.892476: E external/org_tensorflow/tensorflow/core/common_runtime/executor.cc:637] Executor failed to create kernel. Unimplemented: Fusion is not implemented: [BiasAdd,Relu6]
[[{{node Conv/Relu6}}]]
这很奇怪，因为在进行推理时，通常在Docker上安装的tensorflow-gpu都能正常工作。但是，从源代码构建完之后，会出现问题！