1、TensorRT的需要的文件
需要的基本文件(不是必須的)
1>網絡結構文件(deploy.prototxt)
2>訓練的權重模型(net.caffemodel)
TensorRT 2.0 EA版中的sampleMNISTAPI和TensorRT 1.0中的sampleMNISTGIE 幾乎沒有變化,就是不使用caffemodel 文件構建network 的例子。
2、TensorRT支持的層
Convolution: 2D
Activation: ReLU, tanh and sigmoid
Pooling: max and average
ElementWise: sum, product or max of two tensors
LRN: cross-channel only
Fully-connected: with or without bias
SoftMax: cross-channel only
Deconvolution
對於TensorRT 不支持的層,可以先將支持的層跑完,然后將輸出作為caffe的輸入,用caffe再跑,V1不支持TensorRT 和caffe同時工作,V2支持。(例子NVIDIA正在做,后期可能會上傳github)
3、TensorRT 處理流程
基本處理過程:1>caffe model 轉化 gie的model,或者從磁盤或者網絡加載gie可用的model;2>運行GIE引擎(數據提前copy到GPU中);3>提取結果
https://developer.nvidia.com/tensorrt
https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html
https://docs.nvidia.com/deeplearning/sdk/index.html
https://docs.nvidia.com/deeplearning/sdk/tensorrt-developer-guide/index.html#samples
https://www.cnblogs.com/bonelee/p/8311445.html
https://blog.csdn.net/xh_hit/article/details/82917948
https://blog.csdn.net/xh_hit/article/details/82918162
https://github.com/haoxurt/tensorRT_save_serialization