安装
https://docs.openvinotoolkit.org/latest/_docs_install_guides_installing_openvino_linux.html
使用
文档:https://software.intel.com/en-us/openvino-toolkit/documentation/featured
$ --input 一定要给定节点名和端口号
$ python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_model ./recognition_frozen_model.pb --input DatasetToSingleElement:0 --input_shape=[32,32,32,1] --output text_resnet15/conv5_0/relu --log_level=DEBUG
(1)TensorFlow 模型转换成 OpenVINO 的 IR 模型:
https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_Convert_Model_From_TensorFlow.html
(2)OpenVINO 支持的 TensorFlow 的 Layers:
https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_Supported_Frameworks_Layers.html
如果 OpenVINO 中没有 TensorFlow 的 Layers,可以自定义 Layers:
https://docs.openvinotoolkit.org/latest/_docs_HOWTO_Custom_Layers_Guide.html
https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_customize_model_optimizer_Customize_Model_Optimizer.html
目前 OpenVINO 不支持的 TensorFlow Op:
- tf.math.tanh:https://tensorflow.google.cn/api_docs/python/tf/math/tanh
- tf.math.less_equal:https://tensorflow.google.cn/api_docs/python/tf/math/less_equal
- tf.debugging.Assert:https://tensorflow.google.cn/api_docs/python/tf/debugging/Assert
- tf.select:https://stackoverflow.com/questions/41505746/what-is-the-use-of-tf-select
- tf.math.ceil:https://tensorflow.google.cn/api_docs/python/tf/math/ceil
- tf.math.log:https://tensorflow.google.cn/api_docs/python/tf/math/log
- tf.keras.backend.all:https://tensorflow.google.cn/api_docs/python/tf/keras/backend/all
(3)模型转成 IR 模型后,通过 OpenVINO 的 API 实现高效的 inference:
https://docs.openvinotoolkit.org/latest/_docs_IE_DG_Integrate_with_customer_application_new_API.html
(4)自定义模型中的 json + pipeline.config 说明
案例
https://blog.csdn.net/bleedingfight/article/details/86259268