从零开始的目标检测和关键点检测（二）：训练一个Glue的RTMDet模型

从零开始的目标检测和关键点检测（一）：用labelme标注数据集

从零开始的目标检测和关键点检测（三）：训练一个Glue的RTMPose模型

在[1]用labelme标注自己的数据集中已经标注好数据集(关键点和检测框)，通过labelme2coco脚本将所有的labelme json文件集成为两个coco格式的json文件，即train_coco.json和val_coco.json。训练一个RTMDet模型，需要重写config文件。

一、config文件解读

1、数据集类型即coco格式的数据集，metainfo是指框的类别，因为这里只有一个glue的类，因此NUM_CLASSES为1，注意metainfo类别名后的逗号，

# 数据集类型及路径
dataset_type = 'CocoDataset'
data_root = 'data/glue_134_Keypoint/'
metainfo = {'classes': ('glue',)}
NUM_CLASSES = len(metainfo['classes'])

2、加载backnbone预训练权重和RTMDet-tiny预训练权重

# RTMDet-tiny
load_from = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/rtmdet_tiny_8xb32-300e_coco/rtmdet_tiny_8xb32-300e_coco_20220902_112414-78e30dcc.pth'
backbone_pretrain = 'https://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-tiny_imagenet_600e.pth'
deepen_factor = 0.167
widen_factor = 0.375
in_channels = [96, 192, 384]
neck_out_channels = 96
num_csp_blocks = 1
exp_on_reg = False

3、训练参数设置，如epoch、batchsize…

MAX_EPOCHS = 200
TRAIN_BATCH_SIZE = 8
VAL_BATCH_SIZE = 4
stage2_num_epochs = 20
base_lr = 0.004
VAL_INTERVAL = 5  # 每隔多少轮评估保存一次模型权重

4、default_runtime，即默认设置，在config文件夹的default_runtime.py可看到。不同的MM-框架的默认设置不一样(如default_scope = 'mmdet')，可以包含这个.py也可以直接复制过来。

default_scope = 'mmdet'
default_hooks = dict(
    timer=dict(type='IterTimerHook'),
    logger=dict(type='LoggerHook', interval=1),
    param_scheduler=dict(type='ParamSchedulerHook'),
    checkpoint=dict(type='CheckpointHook', interval=10, max_keep_ckpts=2, save_best='coco/bbox_mAP'),
    # auto coco/bbox_mAP_50 coco/bbox_mAP_75 coco/bbox_mAP_s
    sampler_seed=dict(type='DistSamplerSeedHook'),
    visualization=dict(type='DetVisualizationHook'))
env_cfg = dict(
    cudnn_benchmark=False,
    mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
    dist_cfg=dict(backend='nccl'))
vis_backends = [dict(type='LocalVisBackend')]
visualizer = dict(
    type='DetLocalVisualizer',
    vis_backends=[dict(type='LocalVisBackend')],
    name='visualizer')
log_processor = dict(type='LogProcessor', window_size=50, by_epoch=True)
log_level = 'INFO'
load_from = None
resume = False

5、训练超参数配置

train_cfg = dict(
    type='EpochBasedTrainLoop',
    max_epochs=MAX_EPOCHS,
    val_interval=VAL_INTERVAL,
    dynamic_intervals=[(MAX_EPOCHS - stage2_num_epochs, 1)])

val_cfg = dict(type='ValLoop')
test_cfg = dict(type='TestLoop')

# 学习率
param_scheduler = [
    dict(
        type='LinearLR', start_factor=1e-05, by_epoch=False, begin=0,
        end=1000),
    dict(
        type='CosineAnnealingLR',
        eta_min=0.0002,
        begin=150,
        end=300,
        T_max=150,
        by_epoch=True,
        convert_to_iter_based=True)
]

# 优化器
optim_wrapper = dict(
    type='OptimWrapper',
    optimizer=dict(type='AdamW', lr=base_lr, weight_decay=0.05),
    paramwise_cfg=dict(
        norm_decay_mult=0, bias_decay_mult=0, bypass_duplicate=True))
auto_scale_lr = dict(enable=False, base_batch_size=16)

6、数据处理pipeline，做数据预处理(数据增强)

# DataLoader
backend_args = None
train_pipeline = [
    dict(type='LoadImageFromFile', backend_args=None),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(
        type='CachedMosaic',
        img_scale=(640, 640),
        pad_val=114.0,
        max_cached_images=20,
        random_pop=False),
    dict(
        type='RandomResize',
        scale=(1280, 1280),
        ratio_range=(0.5, 2.0),
        keep_ratio=True),
    dict(type='RandomCrop', crop_size=(640, 640)),
    dict(type='YOLOXHSVRandomAug'),
    dict(type='RandomFlip', prob=0.5),
    dict(type='Pad', size=(640, 640), pad_val=dict(img=(114, 114, 114))),
    dict(
        type='CachedMixUp',
        img_scale=(640, 640),
        ratio_range=(1.0, 1.0),
        max_cached_images=10,
        random_pop=False,
        pad_val=(114, 114, 114),
        prob=0.5),
    dict(type='PackDetInputs')
]
test_pipeline = [
    dict(type='LoadImageFromFile', backend_args=None),
    dict(type='Resize', scale=(640, 640), keep_ratio=True),
    dict(type='Pad', size=(640, 640), pad_val=dict(img=(114, 114, 114))),
    dict(
        type='PackDetInputs',
        meta_keys=('img_id', 'img_path', 'ori_shape', 'img_shape',
                   'scale_factor'))
]

7、加载数据和标注并用对应pipeliane做预处理

train_dataloader = dict(
    batch_size=TRAIN_BATCH_SIZE,
    num_workers=4,
    persistent_workers=True,
    sampler=dict(type='DefaultSampler', shuffle=True),
    batch_sampler=None,
    dataset=dict(
        type='CocoDataset',
        data_root=data_root,
        metainfo=metainfo,
        ann_file='train_coco.json',
        data_prefix=dict(img='images/'),
        filter_cfg=dict(filter_empty_gt=True, min_size=32),
        pipeline=train_pipeline,
        backend_args=None),
    pin_memory=True)
val_dataloader = dict(
    batch_size=VAL_BATCH_SIZE,
    num_workers=2,
    persistent_workers=True,
    drop_last=False,
    sampler=dict(type='DefaultSampler', shuffle=False),
    dataset=dict(
        type='CocoDataset',
        data_root=data_root,
        metainfo=metainfo,
        ann_file='val_coco.json',
        data_prefix=dict(img='images/'),
        test_mode=True,
        pipeline=test_pipeline,
        backend_args=None))
test_dataloader = val_dataloader

8、定义模型结构backbone + neck + head

# 模型结构
model = dict(
    type='RTMDet',
    data_preprocessor=dict(
        type='DetDataPreprocessor',
        mean=[103.53, 116.28, 123.675],
        std=[57.375, 57.12, 58.395],
        bgr_to_rgb=False,
        batch_augments=None),
    backbone=dict(
        type='CSPNeXt',
        arch='P5',
        expand_ratio=0.5,
        deepen_factor=deepen_factor,
        widen_factor=widen_factor,
        channel_attention=True,
        norm_cfg=dict(type='SyncBN'),
        act_cfg=dict(type='SiLU', inplace=True),
        init_cfg=dict(
            type='Pretrained',
            prefix='backbone.',
            checkpoint=backbone_pretrain

        )),
    neck=dict(
        type='CSPNeXtPAFPN',
        in_channels=in_channels,
        out_channels=neck_out_channels,
        num_csp_blocks=num_csp_blocks,
        expand_ratio=0.5,
        norm_cfg=dict(type='SyncBN'),
        act_cfg=dict(type='SiLU', inplace=True)),
    bbox_head=dict(
        type='RTMDetSepBNHead',
        num_classes=NUM_CLASSES,
        in_channels=neck_out_channels,
        stacked_convs=2,
        feat_channels=neck_out_channels,
        anchor_generator=dict(
            type='MlvlPointGenerator', offset=0, strides=[8, 16, 32]),
        bbox_coder=dict(type='DistancePointBBoxCoder'),
        loss_cls=dict(
            type='QualityFocalLoss',
            use_sigmoid=True,
            beta=2.0,
            loss_weight=1.0),
        loss_bbox=dict(type='GIoULoss', loss_weight=2.0),
        with_objectness=False,
        exp_on_reg=exp_on_reg,
        share_conv=True,
        pred_kernel_size=1,
        norm_cfg=dict(type='SyncBN'),
        act_cfg=dict(type='SiLU', inplace=True)),
    train_cfg=dict(
        assigner=dict(type='DynamicSoftLabelAssigner', topk=13),
        allowed_border=-1,
        pos_weight=-1,
        debug=False),
    test_cfg=dict(
        nms_pre=30000,
        min_bbox_size=0,
        score_thr=0.001,
        nms=dict(type='nms', iou_threshold=0.65),
        max_per_img=300))

二、开始训练

1、开始训练

python tools/train.py data/glue_134_Keypoint/rtmdet_tiny_glue.py

训练结果

 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.719
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.483
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.766
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.766
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = -1.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = -1.000
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.766

2、测试一下训练结果

python demo/image_demo.py data/glue_134_Keypoint/test_image/test.png data/glue_134_Keypoint/rtmdet_tiny_glue.py --weights work_dirs/rtmdet_tiny_glue/best_coco_bbox_mAP_epoch_180.pth --device cpu

3、可视化训练过程

4、由于标注数据集的glue都是小目标的，因此大目标无法识别，如下：

三、数据集分析

1、可视化部分图像

框标注-框中心点位置分布

框标注-框宽高分布

显然都是小目标的检测

四、ncnn部署

在线模型转换：Deploee

上传文件完成在线转换