"Record" MMDetection Introduction

OpenMMLabis a deep learning high-level library,通过OpenMMLab The software stack can be seen,The lower level is still used Pytorch,So the user is not learning a whole new library,but more advanced,更加方便的Pytorch库,OpenMMLabIt can bring higher efficiency to users,So that users can only focus on the core algorithm,Don't waste more time on configuration files

MMdet 主要由 4 个部分组成,datasets、models、core 和 apis
另外,MMCV is for the whole OpenMMLab All open source libraries,There are general classes and utility classes,Convenient for all downstream codebase 复用

backbone

mmdet/models/backbones

__all__ = [ 'RegNet', 'ResNet', 'ResNetV1d', 'ResNeXt', 'SSDVGG', 'HRNet', 'Res2Net',
		    'HourglassNet', 'DetectoRS_ResNet', 'DetectoRS_ResNeXt', 'Darknet',
		    'ResNeSt', 'TridentResNet'	]

通过 MMCV Registrar mechanism in ,你可以通过 dict form configuration to instantiate any registered class,非常方便和灵活.

# Pretrained weight paths for the skeleton
pretrained='torchvision://resnet50',

backbone=dict(
    type='ResNet', # Skeleton class name,The following parameters are the initialization parameters of the class
    depth=50,
    num_stages=4,
    out_indices=(0, 1, 2, 3),
    frozen_stages=1,
    norm_cfg=dict(type='BN', requires_grad=True), 
    norm_eval=True,
    style='pytorch'),

neck

neck 可以认为是 backbone 和 head 的连接层,主要负责对 backbone The features are efficiently fused and enhanced,The input single-scale or multi-scale features can be fused、Enhanced output, etc.
mmdet/models/necks

__all__ = [
    'FPN', 'BFP', 'ChannelMapper', 'HRFPN', 'NASFPN', 'FPN_CARAFE', 'PAFPN',
    'NASFCOS_FPN', 'RFP', 'YOLOV3Neck'
]

最常用的应该是 FPN,A typical usage is：

neck=dict(
    type='FPN',
    in_channels=[256, 512, 1024, 2048], # 骨架多尺度特征图输出通道
    out_channels=256, # 增强后通道输出
    num_outs=5), # 输出num_outs个多尺度特征图

head

The output of the target detection algorithm generally includes two branches: classification and box coordinate regression,不同算法 head Modules vary in complexity,灵活度比较高.在网络构建方面,To understand the target detection algorithm is mainly to understand head 模块.

MMDetection 中 head Modules are divided into two-stage 所需的 RoIHead 和 one-stage 所需的 DenseHead,

• 所有的 one-stage 算法的 head 模块都在mmdet/models/dense_heads中
• 而 two-stage The algorithm also includes additional mmdet/models/roi_heads

Almost every algorithm includes a standalone head

# dense_heads:
__all__ = [
    'AnchorFreeHead', 'AnchorHead', 'GuidedAnchorHead', 'FeatureAdaption',
    'RPNHead', 'GARPNHead', 'RetinaHead', 'RetinaSepBNHead', 'GARetinaHead',
    'SSDHead', 'FCOSHead', 'RepPointsHead', 'FoveaHead',
    'FreeAnchorRetinaHead', 'ATSSHead', 'FSAFHead', 'NASFCOSHead',
    'PISARetinaHead', 'PISASSDHead', 'GFLHead', 'CornerHead', 'YOLACTHead',
    'YOLACTSegmHead', 'YOLACTProtonet', 'YOLOV3Head', 'PAAHead',
    'SABLRetinaHead', 'CentripetalHead', 'VFNetHead', 'TransformerHead'
]

需要注意的是：two-stage 或者 mutli-stage 算法,Will additionally include a region extractor roi extractor,Used to convert different sizes RoI The feature maps are unified into the same size.

虽然 head Part of the network construction is relatively simple,But due to positive and negative sample property definition、Sum of positive and negative samples bbox 编解码模块都在 head A combined call is made in the module,故
MMDetection The most complex module in head.

下载配置文件

建议在 Conda环境中,在Pycharm中好像不能用,暂时还不知道是什么原因造成的

# 安装 mim
pip install openmim


mim search mmdet --model "mask r-cnn"
mim download mmdet --config "mask_rcnn_r50_fpn_2x_coco" --dest .

注意！--dest . 中间有空格

运行测试模型

from mmdet.apis import init_detector, inference_detector, train_detector, show_result_pyplot
from mmdet.datasets import build_dataset
from mmcv import Config

import mmcv
import os.path as osp
import multiprocessing


if __name__ == '__main__':

    multiprocessing.freeze_support()

    config_file = 'solo_r50_fpn_3x_coco.py'
    checkpoint_file = 'solo_r50_fpn_3x_coco_20210901_012353-11d224d7.pth'

    model = init_detector(config_file, checkpoint_file, device='cuda:0')
    # print(model)

    cfg = Config.fromfile(config_file)

    dataset = [build_dataset(cfg.data.train)]
    mmcv.mkdir_or_exist(osp.abspath('./work_dir_sum'))
    train_detector(model, dataset, cfg, distributed=False, validate=False)


    # 推理演示图像
    result = inference_detector(model, '81.jpg')
    show_result_pyplot(model,'81.jpg', result)