nicklhy / dlinfbench

CNN model inference benchmarks for some popular deep learning frameworks

DLInfBench

Introduction

Benchmarks of the CNN inference task over some popular deep learning frameworks.

Currently, we support five deep learning frameworks: Caffe, Caffe2, PyTorch, MXNet, TensorFlow. Some commonly used imagenet models (i.e. alexnet, resnet50, resnet101 and resnet152) are ready to test. For convenience, we provide all the code or network definition files here. There is no need to download pre-trained weights because we will randomly initialize them.

In order to exclude the impacts of different storage devices and different IO implementations over these deep learning frameworks, we generate all data randomly in advance. The time we calculate in the benchmark experiments only include the cpu-memory to gpu-memory data copy time and the GPU forward time.

I may add benchmark code for more networks (i.e. inception-bn, inception-v3) and deep learning frameworks in the future but no specific plans have been made yet. Thus, anyone is welcomed to submit PRs.

Usage

1. Install Caffe, Caffe2, PyTorch and MXNet in your machine and make sure you can import them in python. Turn on CUDNN support if possible. If you only want to test a part of them, please modify "DLLIB_LIST" in run.sh.
2. Modify the "GPU" variable in run.sh to the gpu device you want to use. (In order to get accurate results, please select a GPU without any other process running on it.)
3. Start benchmark experiments by executing sh run.sh.
4. The results of network's inference speed and gpu memory cost will be saved to cache/results/${DLLIB}_${NETWORK}_${BATCH_SIZE}.txt. Columns in these files represent "framework name", "network", "batch size", "speed(images/s)", "gpu memory(MB)" respectively.
5. We also plot the benchmark results in pictures as it could make them more straightforward. cache/results/${NETWORK}_speed.png demonstrates the network's inference speed of different batch size in different frameworks. cache/results/${NETWORK}_gpu_memory.png demonstrates the network's gpu memory cost of different batch size in different frameworks.

Known Issues

1. There is a problem when I try to run alexnet with CUDNN in caffe2(check the code here). Thus, CUDNN is turned off temporally in caffe2's alexnet benchmarks. If you know how to fix this bug, a PR is welcomed.

Results

• Tesla V100: RESULT_TeslaV100.md.
• Geforce RTX 2080 Ti: RESULT_RTX2080Ti.md.
• TitanV: RESULT_TitanV.md.
• Titan X (Pascal): RESULT_TitanXP.md.
• GeForce GTX 1080 Ti: RESULT_GTX1080Ti.md.
• GeForce GTX 1080: RESULT_GTX1080.md.

License

This project is licensed under an Apache-2.0 license.

dlinfbench's Issues

Re-perform Benchmark with 2022 updated versions?

Hi :)
it is possible to perform the benchmark (with the same hardware or better) with the actual version of DL Frameworks?