EnflameTechnology/hlir_builder_sample

Hlir builder sample

Hlir builder sample

本repo包含了多个常见op和神经网络结构在hlir的实现,用于展示如何基于tops sdk搭建常见op或网络结构的hlir；相关api调用也可参考doc/TopsGraph API参考+用户手册。

安装

本项目运行前请确保安装了tops sdk和驱动

$ dpkg -i tops-sdk.deb

本版本sample在i20的2.0 production(2.0.178)SDK验证可跑通，具体sample执行情况如下：

total_case:43, successed:38, result_failed:0, run_failed:5
successed list:Flatten Dropout ReduceMean Squeeze DotGeneral Slice Sqrt Convert Transpose MaxPool_by_ReduceWindow Reshape Tanh BatchNorm Split AveragePool Add_fp16 Concat Reciprocal Add_fp16_without_convert Clip Tile Copy GlobalAveragePool Add MaxPool Gather ResidualBlock Mul Unsqueeze ResidualBlock_fp16_without_convert Conv2D Sub Relu OneHot Softmax Shape Pow Less
result_failed:
run_failed:Sigmoid Conv Gemm TopK Resize

编译

mkdir build && cd build
cmake .. -DOPType=Add  //构建需要编译的op类型，具体见operator文件夹内
make //编译

目前已实现的op类型：

Add	Add_fp16	Add_fp16_without_convert	AveragePool	BatchNorm
Conv2D	Convert	Copy	DotGeneral	Dropout
Less	MaxPool	GlobalAveragePool	Mul	OneHot
Relu	Reshape	ReduceMean	Squeeze	Resize
Softmax	Split	MaxPool_by_ReduceWindow	Sqrt	Slice
Conv	Gemm	Gather		Tanh
Transpose	Unsqueeze	Sigmoid	Sub	TopK
Concat	Gather	Clip	Flatten	Pow
Shape	Tile

目前已实现的神经网络结构类型：

ResidualBlock	ResidualBlock_fp16_without_convert

or，执行以下脚本可自动验证所有op

./test.sh

TODO

• 完整模型的搭建实例（mnist等）

• 更多网络结构的实例