Detectron 之 config.py 文件参数快速跑自己模型 Mmdetection

「爱情、让人受尽委屈。」 2021-09-17 16:26 615阅读 0赞

![在这里插入图片描述][watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE2MjM0NjEz_size_16_color_FFFFFF_t_70]

## Detectron： ##

[Detectron][]  
1、参数理解  
[中文参数说明][Link 1]  
[官方参数说明][Link 2]  
[多GPU配置][GPU]

2、格式转换  
[训练集格式转换][Link 3]  
[coco格式][coco]  
[roidb格式][roidb]

3、训练流程  
[训练流程1][1]  
[训练流程2][2]

4、其他  
[添加垂直翻转][Link 4]  
[faster rcnn流程：附带部分参数解释][faster rcnn]  
[Faster-RCNN算法精读:最后一段对理解参数有帮助][Faster-RCNN]

## Mmdetection： ##

mmdetection采用注册所有模块，然后使用配置文件构建模块的方式，因此查找整个流程时不易。  
mmdetection数据增强是一次一张图片  
[Mmdetection][]  
mmdetection流程和detectron相似  
[参数解释][Link 5]  
**测试时对增强图片的逆处理流程：**  
[1、增强transform位置][1_transform]  
[2、构建数据集BuildDataset，应用测试增强][2_BuildDataset]  
[3、构建数据集加载器BuildDataLoader][3_BuildDataLoader]  
[4、模型测试流程：是走simple test还是aug test][4_simple test_aug test]  
[5、具体到某个模型Cascade Rcnn][5_Cascade Rcnn]  
[6、simple的逆处理：bbox rescale][6_simple_bbox rescale]  
[7、验证时调用coco接口计算指标][7_coco]

**配置详解：**  
基础参数就不做介绍了

# model settings
    model = dict(
    	# 模块路径：mmdet/models/detectors
        type='CascadeRCNN',
        num_stages=3,
        pretrained='open-mmlab://resnext101_32x4d',
        backbone=dict(
        	# 模块路径： mmdet/models/backbones
            type='ResNeXt',
            depth=101,
            groups=32,
            base_width=4,
            num_stages=4,
            # 除了4个大的stage，底层还有个7*7卷积层
            out_indices=(0, 1, 2, 3),
            # frozen 7*7卷积层 + out_indices为1-1=0的stage
            frozen_stages=1,
            style='pytorch'),
        neck=dict(
        	# 模块路径：mmdet/models/necks
            type='FPN',
            in_channels=[256, 512, 1024, 2048],
            out_channels=256,
            num_outs=5),
        rpn_head=dict(
        	# 模块路径：mmdet/models/anchor_heads
            type='RPNHead',
            in_channels=256,
            feat_channels=256,
            anchor_scales=[8],    # 视数据集目标面积而定
            anchor_ratios=[0.5, 1.0, 2.0],   # 可根据数据集聚类统计
            anchor_strides=[4, 8, 16, 32, 64],  # 框大小 ： (4*8) ** 2
            target_means=[.0, .0, .0, .0],
            target_stds=[1.0, 1.0, 1.0, 1.0],
            loss_cls=dict(
            	# rpn分类只做前背景0-1分类
                type='CrossEntropyLoss', use_sigmoid=True, loss_weight=1.0),  
                # rpn回归只做目标回归，背景0损失 
            loss_bbox=dict(type='SmoothL1Loss', beta=1.0 / 9.0, loss_weight=1.0)),
        bbox_roi_extractor=dict(
        	# 模块路径： mmdet/models/roi_extractors
            type='SingleRoIExtractor',
            roi_layer=dict(type='RoIAlign', out_size=7, sample_num=2),
            out_channels=256,
            featmap_strides=[4, 8, 16, 32]),
        bbox_head=[
        	# 模块路径 ： mmdet/models/bbox_heads
            dict(
                type='SharedFCBBoxHead',
                num_fcs=2,
                in_channels=256,
                fc_out_channels=1024,
                roi_feat_size=7,
                num_classes=81,   # 类别数+1（背景），代码会将anno中的categories按顺序进行映射，因此预测时可能需要反向映射
                target_means=[0., 0., 0., 0.],
                target_stds=[0.1, 0.1, 0.2, 0.2],
                reg_class_agnostic=True,
                loss_cls=dict(
                    type='CrossEntropyLoss', use_sigmoid=False, loss_weight=1.0),
                loss_bbox=dict(type='SmoothL1Loss', beta=1.0, loss_weight=1.0)),
        ],
    # model training and testing settings
    train_cfg = dict(
        rpn=dict(
            assigner=dict(
            	# 模块路径：mmdet/core/bbox/assigners
            	# 注意该模块是在rpn_head.loss里面调用的mmdet/core/anchor/anchor_target.py实现的
            	# rpn的loss训练的bbox没有使用NMS
                type='MaxIoUAssigner',
                pos_iou_thr=0.7,     # 可细致调参
                neg_iou_thr=0.3,
                min_pos_iou=0.3,
                ignore_iof_thr=-1),
            sampler=dict(
            	# 模块路径：mmdet/core/bbox/samplers
            	# 首先预采集256*0.5（不足时就低于）的正样本，然后256-已采集正样本数量得到负样本采集数量进行采集
            	# 对于不足0.5时，可以使用参数neg_pos_ub降低负样本采集数量
                type='RandomSampler',
                num=256,
                pos_fraction=0.5,
                neg_pos_ub=-1,
                add_gt_as_proposals=False),
            allowed_border=0,
            # 可为前背景样本loss赋予权重
            pos_weight=-1,
            debug=False),
        rpn_proposal=dict(
        	# 这一步是一阶段到二阶段的过渡，使用NMS提取兴趣区域
            nms_across_levels=False,
            nms_pre=2000,
            nms_post=2000,
            max_num=2000,
            nms_thr=0.7,    # 可尝试调参
            min_bbox_size=0),
        rcnn=[
            dict(
                assigner=dict(
                	# 该模块在detectors里面调用
                    type='MaxIoUAssigner',
                    pos_iou_thr=0.5,   # 可细致调参
                    neg_iou_thr=0.5,
                    min_pos_iou=0.5,
                    ignore_iof_thr=-1),
                sampler=dict(
                	# 该模块在detectors里面调用
                    type='RandomSampler',
                    num=512,
                    pos_fraction=0.25,
                    neg_pos_ub=-1,
                    add_gt_as_proposals=True),
                pos_weight=-1,    # 可在此处为样本loss加权
                debug=False),
        ],
        stage_loss_weights=[1, 0.5, 0.25])
    test_cfg = dict(
        rpn=dict(
        	# 测试时的兴趣区域提取规则
            nms_across_levels=False,
            nms_pre=1000,
            nms_post=1000,
            max_num=1000,
            nms_thr=0.7,  # 可尝试调参
            min_bbox_size=0),
        rcnn=dict(
        	# 测试时对最终预测的bbos提取规则
        	# bbox分类得分阈值，可细致调参
            score_thr=0.05,
            # NMS的ios阈值，可细致调参
            nms=dict(type='nms', iou_thr=0.5),
            # 单张图片最大输出目标数量
            max_per_img=100))

**其他路径：**  
[验证时加入的hook][hook]  
[数据增强][Link 6]  
[控制样本的哪些信息传入模型][Link 7]  
[anns中categories的顺序与label位置对应][anns_categories_label]

## 目标检测模块结构 ##

[Object Detection and Classification using R-CNNs][]

## 比赛技巧 ##

[Bag of Freebies for Training Object Detection Neural Networks][]  
[目标检测精度提升之奇技淫巧][Link 8]  
[目标检测调优技巧：《Bag of Freebies for Training Object Detection Neural Networks》论文笔记][Bag of Freebies for Training Object Detection Neural Networks 1]  
[我这两年的目标检测][Link 9]  
[目标检测Tricks][Tricks]  
[工业视觉中的目标检测——兼谈天池大赛优胜方案][Link 10]  
[Crowdhuman人体检测比赛第一名经验总结][Crowdhuman]  
[津南数字制造算法挑战赛【赛场二】总决赛 亚军比赛攻略\_这是我的马队][Link 11]  
[目标检测系列三：奇技淫巧][Link 12]

[watermark_type_ZmFuZ3poZW5naGVpdGk_shadow_10_text_aHR0cHM6Ly9ibG9nLmNzZG4ubmV0L3FxXzE2MjM0NjEz_size_16_color_FFFFFF_t_70]: /images/20210520/29a1c6bc155a4124b1996f1af6f923df.png
[Detectron]: https://github.com/facebookresearch/Detectron
[Link 1]: https://blog.csdn.net/zziahgf/article/details/79652946
[Link 2]: https://github.com/facebookresearch/Detectron/blob/master/detectron/core/config.py
[GPU]: https://github.com/facebookresearch/Detectron/blob/master/configs/getting_started/tutorial_1gpu_e2e_faster_rcnn_R-50-FPN.yaml#L14
[Link 3]: https://blog.csdn.net/qq_15969343/article/details/80848175
[coco]: http://cocodataset.org/#format-data
[roidb]: https://www.jianshu.com/p/29ab7457b61e
[1]: https://zhuanlan.zhihu.com/p/50127900
[2]: https://blog.csdn.net/zziahgf/article/details/79488025
[Link 4]: https://www.zybuluo.com/sixijinling/note/1076081
[faster rcnn]: https://www.cnblogs.com/houkai/p/6824455.html
[Faster-RCNN]: https://blog.csdn.net/hunterlew/article/details/71075925
[Mmdetection]: https://github.com/open-mmlab/mmdetection
[Link 5]: https://blog.csdn.net/hajlyx/article/details/85991400
[1_transform]: https://github.com/open-mmlab/mmdetection/blob/ebf59c8958/mmdet/datasets/pipelines/test_aug.py#L8
[2_BuildDataset]: https://github.com/open-mmlab/mmdetection/blob/ebf59c8958/mmdet/datasets/custom.py#L146
[3_BuildDataLoader]: https://github.com/open-mmlab/mmdetection/blob/ebf59c8958/mmdet/datasets/loader/build_loader.py#L61
[4_simple test_aug test]: https://github.com/open-mmlab/mmdetection/blob/b69667001f/mmdet/models/detectors/base.py#L101
[5_Cascade Rcnn]: https://github.com/open-mmlab/mmdetection/blob/b69667001f/mmdet/models/detectors/cascade_rcnn.py#L306
[6_simple_bbox rescale]: https://github.com/open-mmlab/mmdetection/blob/b69667001f/mmdet/models/bbox_heads/bbox_head.py#L137
[7_coco]: https://github.com/open-mmlab/mmdetection/blob/b69667001f/mmdet/core/evaluation/eval_hooks.py#L140
[hook]: https://github.com/open-mmlab/mmdetection/blob/ebf59c8958/mmdet/core/evaluation/eval_hooks.py
[Link 6]: https://github.com/open-mmlab/mmdetection/blob/ebf59c8958/mmdet/datasets/pipelines/transforms.py
[Link 7]: https://github.com/open-mmlab/mmdetection/blob/ebf59c8958/mmdet/datasets/pipelines/formating.py#L136
[anns_categories_label]: https://github.com/open-mmlab/mmdetection/blob/master/mmdet/datasets/coco.py#L29
[Object Detection and Classification using R-CNNs]: http://www.telesens.co/2018/03/11/object-detection-and-classification-using-r-cnns/
[Bag of Freebies for Training Object Detection Neural Networks]: https://arxiv.org/pdf/1902.04103.pdf
[Link 8]: https://blog.csdn.net/qq_14845119/article/details/87379911
[Bag of Freebies for Training Object Detection Neural Networks 1]: https://blog.csdn.net/m_buddy/article/details/88431426
[Link 9]: https://zhuanlan.zhihu.com/p/82491218
[Tricks]: https://zhuanlan.zhihu.com/p/85017172
[Link 10]: https://www.zhuanzhi.ai/document/a09685a02d71b89f26e45be1de7481a9
[Crowdhuman]: https://zhuanlan.zhihu.com/p/68677880
[Link 11]: https://tianchi.aliyun.com/notebook-ai/detail?postId=57797
[Link 12]: http://spytensor.com/index.php/archives/53/