Group20 Resnet18

In this repo, we reimplement ResNet18 with CUDA, and make optimizations to speed up inference. Experiments show that our code is faster than the pytorch version (batchsize=1).

We use im2col and winograd(4x4) to speed up convolution.

onnx模型下载链接：链接：https://pan.baidu.com/s/1eVvb2OedbnYR7m6PG-U_Cw 提取码：ksm8

解析onnx模型并获取参数

1、通过get_onnx_weight.py获得weight.json

cd /home/group20/cuda_onnx_python/
conda activate onnx_env
python get_onnx_weight.py

2、通过Jsoncpp加载参数至模型中，其中/json以及/json_lib即为jsoncpp所需文件

Baseline

通过pytorch搭建Resnet18并将backend设为cudnn作为baseline

cd /home/group20/git/resnet_python/try_resnet_format.py
conda activate onnx_env
python try_resnet_format.py

CUDA搭建与实现

kernels.cu: MaxPooing AvgPooling Relu Add

GEMM: matmul.cu

winograd: conv_winograd_4x4_3x3.cu conv_winograd_gpu.cu

im2col: conv_im2col.cu

resnet_extern.cu: resnet

resnet18_main.cc: main

cd /home/group20/resnet_cuda/tmp/final_version/
make
./hello

输入输出文件及模型

resnet18Input.txt

resnet18Output.txt

resnet18.onnx

weight.json

实验结果

Methods	time(ms)
Baseline (pytorch)	2.67
our model	2.26