Mmdet benchmark with python

Overview

mmdet_benchmark

本项目是为了研究 mmdet 推断性能瓶颈,并且对其进行优化。

配置与环境

机器配置

CPU:Intel(R) Core(TM) i9-10900K CPU @ 3.70GHz
GPU:NVIDIA GeForce RTX 3080 10GB
内存:64G
硬盘:1TB NVME SSD

mmdet 环境

Python: 3.9.7 (default, Sep 16 2021, 13:09:58) [GCC 7.5.0]
CUDA available: True
GPU 0: NVIDIA GeForce RTX 3080
CUDA_HOME: /usr/local/cuda
NVCC: Build cuda_10.2_r440.TC440_70.29663091_0
GCC: gcc (Ubuntu 7.5.0-3ubuntu1~18.04) 7.5.0
PyTorch: 1.9.1+cu111
PyTorch compiling details: PyTorch built with:
  - GCC 7.3
  - C++ Version: 201402
  - Intel(R) Math Kernel Library Version 2020.0.0 Product Build 20191122 for Intel(R) 64 architecture applications
  - Intel(R) MKL-DNN v2.1.2 (Git Hash 98be7e8afa711dc9b66c8ff3504129cb82013cdb)
  - OpenMP 201511 (a.k.a. OpenMP 4.5)
  - NNPACK is enabled
  - CPU capability usage: AVX2
  - CUDA Runtime 11.1
  - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86
  - CuDNN 8.0.5
  - Magma 2.5.2
  - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.1, CUDNN_VERSION=8.0.5, CXX_COMPILER=/opt/rh/devtoolset-7/root/usr/bin/c++, CXX_FLAGS= -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DNDEBUG -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_VERSION=1.9.1, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON,

TorchVision: 0.10.1+cu111
OpenCV: 4.5.4
MMCV: 1.3.17
MMCV Compiler: GCC 7.3
MMCV CUDA Compiler: 11.1
MMDetection: 2.19.0+

时间分析

Mask R-CNN 的推断过程包含以下几个步骤,我们在一些可能是瓶颈的位置增加了时间统计:

注意:mask post-processing 的时间包含在 roi_head 里,所以减少 mask post-processing 的时间就是在减少 roi_head 的时间。

使用标准尺寸测试(1333x800)

测试图片:

stage pre-processing backbone rpn_head mask forward mask post-processing roi_head total
inference 13.45 24.87 10.16 3.84 15.74 23.49 72.3
inference_fp16 13.53 15.98 8.34 3.36 15.74 22.97 61.4
inference_fp16_preprocess 1.75 15.91 8.21 3.33 15.61 22.69 49.03
inference_raw_mask 1.65 15.93 8.34 3.36 1.74 8.89 33.45

使用较大尺寸测试(3840x2304)

stage pre-processing backbone rpn_head mask forward mask post-processing roi_head total
inference 128.44 187.24 69.96 6.01 173.72 183.95 569.92
inference_fp16 127.28 120.10 50.30 6.80 172.42 186.81 485.04
inference_fp16_preprocess 11.02 120.20 50.18 6.82 174.62 187.07 379.00
inference_raw_mask 11.03 120.26 50.46 6.81 2.99 15.34 197.69

可视化

mmdet 原版:

加速版:

目测没有显著差异。

总结

  • 使用 wrap_fp16_model 可以节省 backbone 的时间,但是不是所有情况下的 forward 都能节省时间;
  • 使用 torchvision.transforms.functional 去做图像预处理,可以极大提升推断速度;
  • 使用 FCNMaskHeadWithRawMask,避免对 mask 进行 resize,对越大的图像加速比越高,因为 resize 到原图大小的成本很高;
  • 后续优化,需要考虑 backbonerpn_head 的优化,可以使用 TensorRT 进行加速。

原理分析

fp16

把一些支持 fp16 的层使用 fp16 来推断,可以充分利用显卡的 TensorCore,加速 forward 部分的速度。

参考链接:https://zhuanlan.zhihu.com/p/375224982

在 backbone 上,时间从 24.87 降到 15.93,在大图上从 187.24 降到 120.26,提升 35% 左右。

torchvision.transforms.functional

使用 pytorch 的 resize、pad、normalize,可以利用上 GPU 而不是 CPU。我们在推断过程中,CPU 利用率始终是最高的,而 GPU 利用率几乎没有满过,所以只要能够把 CPU 的事情交给 GPU 做,就能解决瓶颈问题,减少推断时间。

由于整个过程都可以使用 GPU,所以时间从 13.45 降低到 1.65,在大图上从 128.44 降低到 11.03,提升 10 倍左右。

FCNMaskHeadWithRawMask

首先我们看看 mmdet 处理的结果格式:

可以看到,有多少个 bbox,就有多少个 segm,每个 segm 都是原图尺寸。不管是 CPU,还是内存,都需要大量的时间去处理。

然后再看看 FCNMaskHeadWithRawMask 处理的格式:

每个结果都是 28x28 的,这也是模型原始输出,所以信息量和上面是一样的。

唯一的区别是,我们在拿到结果之后,如果要可视化,需要 resize 到 bbox 的大小,参考 detect/utils_visualize.py#L36-L40

使用 FCNMaskHeadWithRawMask 可以从 15.74 降到 1.74,大图可以从 173.72 降到 2.99,也就是说,图越大,这个加速比越大。

You might also like...
OpenMMLab Semantic Segmentation Toolbox and Benchmark.
OpenMMLab Semantic Segmentation Toolbox and Benchmark.

Documentation: https://mmsegmentation.readthedocs.io/ English | 简体中文 Introduction MMSegmentation is an open source semantic segmentation toolbox based

OpenMMLab's Next Generation Video Understanding Toolbox and Benchmark
OpenMMLab's Next Generation Video Understanding Toolbox and Benchmark

Introduction English | 简体中文 MMAction2 is an open-source toolbox for video understanding based on PyTorch. It is a part of the OpenMMLab project. The m

OpenMMLab Pose Estimation Toolbox and Benchmark.
OpenMMLab Pose Estimation Toolbox and Benchmark.

Introduction English | 简体中文 MMPose is an open-source toolbox for pose estimation based on PyTorch. It is a part of the OpenMMLab project. The master b

[CVPR 2021 Oral] ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis
[CVPR 2021 Oral] ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis

ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis ForgeryNet: A Versatile Benchmark for Comprehensive Forgery Analysis [arxiv|pdf|v

A Research-oriented Federated Learning Library and Benchmark Platform for Graph Neural Networks. Accepted to ICLR'2021 - DPML and MLSys'21 - GNNSys workshops.

FedGraphNN: A Federated Learning System and Benchmark for Graph Neural Networks A Research-oriented Federated Learning Library and Benchmark Platform

RoboDesk A Multi-Task Reinforcement Learning Benchmark
RoboDesk A Multi-Task Reinforcement Learning Benchmark

RoboDesk A Multi-Task Reinforcement Learning Benchmark If you find this open source release useful, please reference in your paper: @misc{kannan2021ro

[CVPR2021] UAV-Human: A Large Benchmark for Human Behavior Understanding with Unmanned Aerial Vehicles

UAV-Human Official repository for CVPR2021: UAV-Human: A Large Benchmark for Human Behavior Understanding with Unmanned Aerial Vehicle Paper arXiv Res

3D AffordanceNet is a 3D point cloud benchmark consisting of 23k shapes from 23 semantic object categories, annotated with 56k affordance annotations and covering 18 visual affordance categories.
3D AffordanceNet is a 3D point cloud benchmark consisting of 23k shapes from 23 semantic object categories, annotated with 56k affordance annotations and covering 18 visual affordance categories.

3D AffordanceNet This repository is the official experiment implementation of 3D AffordanceNet benchmark. 3D AffordanceNet is a 3D point cloud benchma

NAS Benchmark in
NAS Benchmark in "Prioritized Architecture Sampling with Monto-Carlo Tree Search", CVPR2021

NAS-Bench-Macro This repository includes the benchmark and code for NAS-Bench-Macro in paper "Prioritized Architecture Sampling with Monto-Carlo Tree

Releases(v0.2.1)
Owner
杨培文 (Yang Peiwen)
杨培文 (Yang Peiwen)
KGDet: Keypoint-Guided Fashion Detection (AAAI 2021)

KGDet: Keypoint-Guided Fashion Detection (AAAI 2021) This is an official implementation of the AAAI-2021 paper "KGDet: Keypoint-Guided Fashion Detecti

Qian Shenhan 35 Dec 29, 2022
Learning Neural Network Subspaces

Learning Neural Network Subspaces Welcome to the codebase for Learning Neural Network Subspaces by Mitchell Wortsman, Maxwell Horton, Carlos Guestrin,

Apple 117 Nov 17, 2022
Library to enable Bayesian active learning in your research or labeling work.

Bayesian Active Learning (BaaL) BaaL is an active learning library developed at ElementAI. This repository contains techniques and reusable components

ElementAI 687 Dec 25, 2022
Creating a Linear Program Solver by Implementing the Simplex Method in Python with NumPy

Creating a Linear Program Solver by Implementing the Simplex Method in Python with NumPy Simplex Algorithm is a popular algorithm for linear programmi

Reda BELHAJ 2 Oct 12, 2022
(AAAI 2021) Progressive One-shot Human Parsing

End-to-end One-shot Human Parsing This is the official repository for our two papers: Progressive One-shot Human Parsing (AAAI 2021) End-to-end One-sh

54 Dec 30, 2022
Code for HodgeNet: Learning Spectral Geometry on Triangle Meshes, in SIGGRAPH 2021.

HodgeNet | Webpage | Paper | Video HodgeNet: Learning Spectral Geometry on Triangle Meshes Dmitriy Smirnov, Justin Solomon SIGGRAPH 2021 Set-up To ins

Dima Smirnov 61 Nov 27, 2022
The official implementation of Autoregressive Image Generation using Residual Quantization (CVPR '22)

Autoregressive Image Generation using Residual Quantization (CVPR 2022) The official implementation of "Autoregressive Image Generation using Residual

Kakao Brain 529 Dec 30, 2022
Dynamic Divide-and-Conquer Adversarial Training for Robust Semantic Segmentation (ICCV2021)

Dynamic Divide-and-Conquer Adversarial Training for Robust Semantic Segmentation This is a pytorch project for the paper Dynamic Divide-and-Conquer Ad

DV Lab 29 Nov 21, 2022
This is the PyTorch implementation of GANs N’ Roses: Stable, Controllable, Diverse Image to Image Translation

Official PyTorch repo for GAN's N' Roses. Diverse im2im and vid2vid selfie to anime translation.

1.1k Jan 01, 2023
FeTaQA: Free-form Table Question Answering

FeTaQA: Free-form Table Question Answering FeTaQA is a Free-form Table Question Answering dataset with 10K Wikipedia-based {table, question, free-form

Language, Information, and Learning at Yale 40 Dec 13, 2022
Semiconductor Machine learning project

Wafer Fault Detection Problem Statement: Wafer (In electronics), also called a slice or substrate, is a thin slice of semiconductor, such as a crystal

kunal suryawanshi 1 Jan 15, 2022
Learning and Building Convolutional Neural Networks using PyTorch

Image Classification Using Deep Learning Learning and Building Convolutional Neural Networks using PyTorch. Models, selected are based on number of ci

Mayur 126 Dec 22, 2022
Pytorch reimplementation of PSM-Net: "Pyramid Stereo Matching Network"

This is a Pytorch Lightning version PSMNet which is based on JiaRenChang/PSMNet. use python main.py to start training. PSM-Net Pytorch reimplementatio

XIAOTIAN LIU 1 Nov 25, 2021
Jaxtorch (a jax nn library)

Jaxtorch (a jax nn library) This is my jax based nn library. I created this because I was annoyed by the complexity and 'magic'-ness of the popular ja

nshepperd 17 Dec 08, 2022
Rotation Robust Descriptors

RoRD Rotation-Robust Descriptors and Orthographic Views for Local Feature Matching Project Page | Paper link Evaluation and Datasets MMA : Training on

Udit Singh Parihar 25 Nov 15, 2022
A python program to hack instagram

hackinsta a program to hack instagram Yokoback_(instahack) is the file to open, you need libraries write on import. You run that file in the same fold

2 Jan 22, 2022
MobileNetV1-V2,MobileNeXt,GhostNet,AdderNet,ShuffleNetV1-V2,Mobile+ViT etc.

MobileNetV1-V2,MobileNeXt,GhostNet,AdderNet,ShuffleNetV1-V2,Mobile+ViT etc. ⭐⭐⭐⭐⭐

568 Jan 04, 2023
A Simple Long-Tailed Rocognition Baseline via Vision-Language Model

BALLAD This is the official code repository for A Simple Long-Tailed Rocognition Baseline via Vision-Language Model. Requirements Python3 Pytorch(1.7.

Teli Ma 4 Jan 20, 2022
PantheonRL is a package for training and testing multi-agent reinforcement learning environments.

PantheonRL is a package for training and testing multi-agent reinforcement learning environments. PantheonRL supports cross-play, fine-tuning, ad-hoc coordination, and more.

Stanford Intelligent and Interactive Autonomous Systems Group 57 Dec 28, 2022
Offline Reinforcement Learning with Implicit Q-Learning

Offline Reinforcement Learning with Implicit Q-Learning This repository contains the official implementation of Offline Reinforcement Learning with Im

Ilya Kostrikov 125 Dec 31, 2022