TNN notes

(TNN frame ）

Model deployment ：

1. Model transformation ：

Will be in tensorflow, pytorch, caffe Wait for the model trained by the platform , First turn to unified onnx Format , Re convert to TNN Format model .

tnn The environment that model transformation needs to rely on is relatively complex , It is suggested to directly use the official built docker.

docker pull turandotkay/tnn-convert:latest

To validate the docker Whether the image can be used normally ：

docker run  -it tnn-convert:latest  python3 ./converter.py -h

At present, we support onnx2tnn, caffe2tnn, tf2tnn, tflite2tnn

usage: python3 ./converter.py onnx2tnn -tp ONNX_PATH -in input_name -on output_name [-o OUTPUT_DIR] [-v v1.0] [-optimize] [-half] [-align]

v： Specify the version number of the model , In order to track and distinguish the model later .

output_dir: This parameter is generally not used , Pair will be generated by default TNN The model is placed in the same path as the current model to be converted .

optimize: Optimize the model , Strongly enable this option , Only when turning on this option fails to convert the model , It is recommended to remove this option and try again .

half: The model data passes through FP16 For storage , Reduce the size of the model . Default by FP32 Storage model .

align: Will be converted to TNN Align the model with the original model , determine TNN Whether the model is successfully converted . Currently, only single input single output model and single input multiple output model are supported .align Only support FP32 Model verification .

So use align You can't use half.

input_file: Specify the input file name required for model alignment .

ref_file: Specify the name of the output file to be aligned .

example: docker run --volume=${pwd}:/workspace -it tnn-convert:latest python3 ./converter.py onnx2tnn \

/workspace/mobilenetv3-small-c7eb32fe.onnx \

-optimize \

-v v3.0 \

-align 

2. Build engine ：

Hardware support for the target platform , Complete the target platform TNN Engine compilation .

Mobile demo：

1. iOS demo

2. Android demo