mmdet实验步骤以及常见报错应对

配环境时用pytorch2.0以上好些。一次成功见这个教程

具体布置自己的东西时，可以参考这个教程

数据集准备

数据如下准备。然后test2017用不着，空着

跑实验

1/选择一个模型 使用python tools/train.py configs/...运行，找到work dirs目录下的模型py文件

2/修改配置文件训练测测试相关参数:修改检测数，一般是classes

3/重载数据集的类名。去官网下载权重，把配置文件中load_from = none 修改为load_form = '权重路径'

4/运行修改好后的配置文件 python tools/train.py work dirs/...xxx.py，得到相应结果

8/ 运行python tools/  workdirs/....py 来测试模型的参数大小，flops和fps

常见报错

1．File "<__array_function__ internals>", line 200, in concatenate ValueError: need at least one array to concatenate

解决：1.在配置文件的三个datalodaer里面，data_root='data/coco/',后面加上metainfo

2.确保classes设置的数量对了

如metainfo=dict(

            classes=(

                'Yagi_antenna',

                'Plate_log_antenna',

                'Patch_antenna',

            ),

            palette=[

                (

                    220,

                    20,

                    60,

                ),

                (

                    119,

                    11,

                    32,

                ),

                (

                    0,

                    0,

                    142,

                ),

            ]),

变成这样：

train_dataloader = dict(

    batch_sampler=dict(_scope_='mmdet', type='AspectRatioBatchSampler'),

    batch_size=2,

    dataset=dict(

        _scope_='mmdet',

        ann_file='annotations/instances_train2017.json',

        backend_args=None,

        data_prefix=dict(img='train2017/'),

        data_root='data/coco/',

        filter_cfg=dict(filter_empty_gt=True, min_size=32),

        metainfo=dict(

            classes=(

                'Yagi_antenna',

                'Plate_log_antenna',

                'Patch_antenna',

            ),

            palette=[

                (

                    220,

                    20,

                    60,

                ),

                (

                    119,

                    11,

                    32,

                ),

                (

                    0,

                    0,

                    142,

                ),

            ]),

        pipeline=[

            dict(backend_args=None, type='LoadImageFromFile'),

            dict(type='LoadAnnotations', with_bbox=True, with_mask=True),

            dict(

                keep_ratio=True,

                ratio_range=(

                    0.1,

                    2.0,

                ),

                scale=(

                    1024,

                    1024,

                ),

                type='RandomResize'),

            dict(

                allow_negative_crop=True,

                crop_size=(

                    1024,

                    1024,

                ),

                crop_type='absolute_range',

                recompute_bbox=True,

                type='RandomCrop'),

            dict(min_gt_bbox_wh=(

                0.01,

                0.01,

            ), type='FilterAnnotations'),

            dict(prob=0.5, type='RandomFlip'),

            dict(type='PackDetInputs'),

        ],

        type='CocoDataset'),

    num_workers=4,

    persistent_workers=True,

    sampler=dict(_scope_='mmdet', shuffle=True, type='DefaultSampler'))

2.data'category id']= self.cat ids[label] IndexError: list index out of range 

参考这个教程。很有用，我是当时代码中的类名由a-b 变成了a - b，中间多了空格，导致代码和json里面的标注不一样而报错。

3.you should set `PYTHONPATH` to make 'sys.path include the directory which contains your custom module

解决：import sys

sys.path.append("/hy-tmp/mmdetection")