[Report]
Authors: Hongjie Fang, Peishen Yan, Haoran Zhao.
This is the inference booster of the InceptionV3[1] model. Features includes:
- Implementation of convolution in CPU, CUDA, CUDNN.
- Optimization of convolution (implicit im2col and tilling method).
- Implementation of pooling and FC layer in CPU, CUDA, CUDNN.
- Optimization of the FC layer using tilling method.
- Implementation of the full Inception-v3 network in CPU, CUDA and CUDNN.
- Pytorch inference implementation[2] of Inception-v3 network (only for debug use).
- ONNX-to-JSON formatter for Inception-v3 onnx model.
This is also the final project of course "CS433: Parallel and Distributed Computing" of Shanghai Jiao Tong University, taught by Prof. Xiaoyao Liang.
Compile the source codes.
cd src
make
cd ..You may need to change the nvcc path in src/makefile. Different compile options are required for different architecture. We only provide compile options for our experiment architecture (Tesla V100, CUDA 10.2).
Download data from Baidu Netdisk (Verify code: csov), and put it in the data folder under the root directory of the repository. Then, you can test the inception code using the given model, input and output.
cd test
./inception_main
cd ..The experiment will run for approximately 10 minutes, which includes 5,000 inference experiments. Here are some experiment statistics.
| Implementation method | Average Inference Time | Max GPU occupation |
|---|---|---|
| CPU | ~180,000 ms | - |
| Our basic CUDA Implementaion | ~36,000 ms | 530 MB |
| CUDNN | 102.594 ms | 750 MB |
| Our CUDA Implementation | 61.096 ms | 530 MB |
The result show that our implementation is faster than the default implementation of CUDNN.
 |  |
| Test result of our implementations | Test result of our CUDNN implementations |
[1] Szegedy, Christian, et al. "Rethinking the inception architecture for computer vision." Proceedings of the IEEE conference on computer vision and pattern recognition. 2016;
[2] https://github.com/zt1112/pytorch_inceptionv3/blob/master/inception3.py.