MMDetection框架训练、测试全流程

前言

• MMDetection是一个目标检测工具箱，包含了丰富的目标检测、实例分割、全景分割算法以及相关的组件和模块，github项目地址。
• 支持的目标检测（Object Detection）模型（近年来的一些SOTA模型）：DAB-DETR、RTMDet、GLIP、Detic、DINO
• 支持的实例分割（Instance Segmentation）模型（近年来的一些SOTA模型）：Mask2former、BoxInst、SparseInst、RTMDet
• 支持的全景分割（Panoptic Segmentation）模型：Panoptic FPN、MaskFormer、Mask2Former
• 关于实例分割和全景分割的区别：全景分割同时提供了像素级别的语义类别和实例标识符，而实例分割只关注物体实例的边界和分割。全景分割提供了更全面的信息，适用于需要对每个像素进行细粒度分析的任务，如自动驾驶。实例分割更专注于检测和分割物体实例，适用于目标检测和图像分割等任务。
• 本文主要介绍了MMDetection的训练与测试过程，在数据集Dog and Cat Detection上微调了RTMDet模型，解析了RTMDet模型，最终模型指标bbox_mAP达到了0.952。

环境配置

• 完整的环境配置代码如下，如果不想看分步解析可以直接跳过本节剩余的内容：

import IPython.display as display

!pip install openmim
!mim install mmengine==0.7.2
# 构建wheel，需要30分钟，构建好以后将whl文件放入单独的文件夹
# !git clone https://github.com/open-mmlab/mmcv.git
# !cd mmcv && CUDA_HOME=/usr/local/cuda-11.8 MMCV_WITH_OPS=1 pip wheel --wheel-dir=/kaggle/working .
!pip install -q /kaggle/input/frozen-packages-mmdetection/mmcv-2.0.1-cp310-cp310-linux_x86_64.whl

!rm -rf mmdetection
!git clone https://github.com/open-mmlab/mmdetection.git
!git clone https://github.com/open-mmlab/mmyolo.git
%cd mmdetection

%pip install -e .

!pip install wandb
display.clear_output()

• 首先安装open-mmlab的包管理库openmim，然后安装mmengine库，代码如下：

!pip install openmim
!mim install mmengine==0.7.2

• 由于在kaggle中无法通过mim直接安装mmcv（后续训练会报错）,我们只能通过构建wheel的方式安装，代码如下：

!git clone https://github.com/open-mmlab/mmcv.git
!cd mmcv && CUDA_HOME=/usr/local/cuda-11.8 MMCV_WITH_OPS=1 pip wheel --wheel-dir=/kaggle/working .

• 上面一步需要等待大概30分钟的时间，然后你就会在/kaggle/working目录下发现mmcv-2.0.1-cp310-cp310-linux_x86_64.whl文件，使用pip install -q /kaggle/working/mmcv-2.0.1-cp310-cp310-linux_x86_64.whl安装即可。但为了节省时间，防止每次运行都需要等很长时间，我将构建的wheel下载然后上传到kaggle Datasets这样每次只用加载数据集就可以安装了，这里提供数据地址。所以安装代码变为：

!pip install -q /kaggle/input/frozen-packages-mmdetection/mmcv-2.0.1-cp310-cp310-linux_x86_64.whl

• 通过git clone的方式安装mmdetection，因为数据集为.xml后缀，后面我们需要使用mmyolo中的工具转换格式，所以一起下载，但不安装mmyolo。

!rm -rf mmdetection
!git clone https://github.com/open-mmlab/mmdetection.git
!git clone https://github.com/open-mmlab/mmyolo.git

# 进入mmdetection项目文件夹
%cd mmdetection

# 安装mmdetection
%pip install -e .

• 如果安装过程中出现pycocotools安装问题，可以参考我的上一篇文章MMYOLO框架标注、训练、测试全流程（补充篇），里面有详细的解决方案。
• 因为在训练过程中需要可视化各项指标，所以安装wandb包，并登录。

!pip install wandb

import wandb
wandb.login()

模型推理

• 我们首先创建一个文件夹checkpoints，用于存放模型的预训练权重。因为我们选择的是RTMDet模型，所以下载对应权重。
• 我们可以打开mmdetection的github项目地址，进入configs/rtmdet路径，在README.md文件中有详细的预训练权重。
  
• 可以看到，模型参数量（Params）越多，精度指标（box AP）越高，我们选择一个参数量适中的模型RTMDet-l，对应的configs文件名为rtmdet_l_8xb32-300e_coco.py。意思是RTMDet-l型号，在8个GPU上，每个GPUbatch size为32，在coco数据集上训练了300epochs的权重。下载并保存在checkpoints文件夹下

!mkdir ./checkpoints
!mim download mmdet --config rtmdet_l_8xb32-300e_coco --dest ./checkpoints

• 使用模型进行推理，并可视化推理结果

from mmdet.apis import DetInferencer

model_name = 'rtmdet_l_8xb32-300e_coco'
checkpoint = './checkpoints/rtmdet_l_8xb32-300e_coco_20220719_112030-5a0be7c4.pth'

device = 'cuda:0'

inferencer = DetInferencer(model_name, checkpoint, device)

img = './demo/demo.jpg'

result = inferencer(img, out_dir='./output')
display.clear_output()

from PIL import Image
Image.open('./output/vis/demo.jpg')

• 如果到这里都没有出现任何问题，说明环境配置的非常成功，RTMDet模型做出了推理。

数据整理

• 数据集Dog and Cat Detection文件组织信息：

 - Dog-and-Cat-Detection
     - annotations
         - Cats_Test0.xml
         - Cats_Test1.xml
         - Cats_Test2.xml
         - ...
     - images
         - Cats_Test0.png
         - Cats_Test1.png
         - Cats_Test2.png
         - ...

• 由于kaggle中在input路径下的数据集是只读类型，不允许更改，并且标注文件为.xml格式，需要转换，这里先将图片复制到./data/images目录下

import shutil

# 复制文件到工作目录
shutil.copytree('/kaggle/input/dog-and-cat-detection/images', './data/images')

• 由于后续切分数据集需要标注信息为.json格式，我们将dog-and-cat-detection/annotations文件夹中的.xml文件转换为1个.json文件。

import xml.etree.ElementTree as ET
import os
import json

coco = dict()
coco['images'] = []
coco['type'] = 'instances'
coco['annotations'] = []
coco['categories'] = []

category_set = dict()
image_set = set()

category_item_id = -1
image_id = 0
annotation_id = 0


def addCatItem(name):
    global category_item_id
    category_item = dict()
    category_item['supercategory'] = 'none'
    category_item_id += 1
    category_item['id'] = category_item_id
    category_item['name'] = name
    coco['categories'].append(category_item)
    category_set[name] = category_item_id
    return category_item_id


def addImgItem(file_name, size):
    global image_id
    if file_name is None:
        raise Exception('Could not find filename tag in xml file.')
    if size['width'] is None:
        raise Exception('Could not find width tag in xml file.')
    if size['height'] is None:
        raise Exception('Could not find height tag in xml file.')
    image_id += 1
    image_item = dict()
    image_item['id'] = image_id
    image_item['file_name'] = file_name + ".png"
    image_item['width'] = size['width']
    image_item['height'] = size['height']
    coco['images'].append(image_item)
    image_set.add(file_name)
    return image_id


def addAnnoItem(object_name, image_id, category_id, bbox):
    global annotation_id
    annotation_item = dict()
    annotation_item['segmentation'] = []
    seg = []
    seg.append(bbox[0])
    seg.append(bbox[1])
    seg.append(bbox[0])
    seg.append(bbox[1] + bbox[3])
    seg.append(bbox[0] + bbox[2])
    seg.append(bbox[1] + bbox[3])
    seg.append(bbox[0] + bbox[2])
    seg.append(bbox[1])

    annotation_item['segmentation'].append(seg)

    annotation_item['area'] = bbox[2] * bbox[3]
    annotation_item['iscrowd'] = 0
    annotation_item['ignore'] = 0
    annotation_item['image_id'] = image_id
    annotation_item['bbox'] = bbox
    annotation_item['category_id'] = category_id
    annotation_id += 1
    annotation_item['id'] = annotation_id
    coco['annotations'].append(annotation_item)


def parseXmlFiles(xml_path):
    for f in os.listdir(xml_path):
        if not f.endswith('.xml'):
            continue
        xmlname = f.split('.xml')[0]

        bndbox = dict()
        size = dict()
        current_image_id = None
        current_category_id = None
        file_name = None
        size['width'] = None
        size['height'] = None
        size['depth'] = None

        xml_file = os.path.join(xml_path, f)

        tree = ET.parse(xml_file)
        root = tree.getroot()
        if root.tag != 'annotation':
            raise Exception('pascal voc xml root element should be annotation, rather than {}'.format(root.tag))

        for elem in root:
            current_parent = elem.tag
            current_sub = None
            object_name = None

            if elem.tag == 'folder':
                continue

            if elem.tag == 'filename':
                file_name = xmlname
                if file_name in category_set:
                    raise Exception('file_name duplicated')

            elif current_image_id is None and file_name is not None and size['width'] is not None:
                if file_name not in image_set:
                    current_image_id = addImgItem(file_name, size)
                else:

                    raise Exception('duplicated image: {}'.format(file_name))

            for subelem in elem:
                bndbox['xmin'] = None
                bndbox['xmax'] = None
                bndbox['ymin'] = None
                bndbox['ymax'] = None

                current_sub = subelem.tag
                if current_parent == 'object' and subelem.tag == 'name':
                    object_name = subelem.text
                    if object_name not in category_set:
                        current_category_id = addCatItem(object_name)
                    else:
                        current_category_id = category_set[object_name]

                elif current_parent == 'size':
                    if size[subelem.tag] is not None:
                        raise Exception('xml structure broken at size tag.')
                    size[subelem.tag] = int(subelem.text)

                for option in subelem:
                    if current_sub == 'bndbox':
                        if bndbox[option.tag] is not None:
                            raise Exception('xml structure corrupted at bndbox tag.')
                        bndbox[option.tag] = int(float(option.text))

                if bndbox['xmin'] is not None:
                    if object_name is None:
                        raise Exception('xml structure broken at bndbox tag')
                    if current_image_id is None:
                        raise Exception('xml structure broken at bndbox tag')
                    if current_category_id is None:
                        raise Exception('xml structure broken at bndbox tag')
                    bbox = []
                    bbox.append(bndbox['xmin'])
                    bbox.append(bndbox['ymin'])
                    bbox.append(bndbox['xmax'] - bndbox['xmin'])
                    bbox.append(bndbox['ymax'] - bndbox['ymin'])
                    addAnnoItem(object_name, current_image_id, current_category_id, bbox)

os.makedirs('./data/annotations')
xml_path = '/kaggle/input/dog-and-cat-detection/annotations'
json_file = './data/annotations/annotations_all.json'
parseXmlFiles(xml_path)
json.dump(coco, open(json_file, 'w'))

• 当前工作目录数据存储文件组织信息：

 - mmdetection
    - data
         - annotations
             - annotations_all.json
         - images
             - Cats_Test0.png
             - Cats_Test1.png
             - Cats_Test2.png
             - ....
     - ...

• 由于我们需要使用mmyolo项目文件中的一个脚本，将数据分为训练和测试集，先进入mmyolo项目文件夹

# 切换到mmyolo项目文件夹
%cd /kaggle/working/mmyolo

• 切分脚本文件位于tools/misc/coco_split.py，参数由上到下分别为： --json（生成的.json文件路径）；–out-dir（生成的切分.json文件存储文件夹路径）；–ratios 0.8 0.2（训练集、测试集占比）；–shuffle（是否打乱顺序）；–seed（随机数种子）

# 切分训练、测试集
!python tools/misc/coco_split.py --json /kaggle/working/mmdetection/data/annotations/annotations_all.json \
                                --out-dir /kaggle/working/mmdetection/data/annotations \
                                --ratios 0.8 0.2 \
                                --shuffle \
                                --seed 2023

• 输出：

Split info: ====== 
Train ratio = 0.8, number = 2949
Val ratio = 0, number = 0
Test ratio = 0.2, number = 737
Set the global seed: 2023
shuffle dataset.
Saving json to /kaggle/working/mmdetection/data/annotations/trainval.json
Saving json to /kaggle/working/mmdetection/data/annotations/test.json
All done!

• 接着切换回mmdetection项目文件夹：

%cd /kaggle/working/mmdetection

• 此时工作目录数据存储文件组织信息：

 - mmdetection
    - data
         - annotations
             - test.json
             - trainval.json
             - annotations_all.json
         - images
             - Cats_Test0.png
             - Cats_Test1.png
             - Cats_Test2.png
             - ....
     - ...

编辑RTMDet模型配置

• RTMDet模型架构图可以在对应参数文件夹README.md文档中找到。
  

• 可以在github中打开configs/rtmdet/rtmdet_l_8xb32-300e_coco.py配置文件（观察_base_值，若有继承关系，可以一直往上查找，直到找到主文件），这里RTMDet-l型号模型已经是主文件了，可以直接查看。

• 我们要更改的主要就是_base_（继承的上级文件）、data_root（数据存储的文件夹）、train_batch_size_per_gpu（每个GPU训练的batch size）、train_num_workers（核心工作数，一般为n GPU x 4）、max_epochs（最大epoch数）、base_lr（基础学习率）、metainfo（种类信息及各种类对应调色板）、train_dataloader（图片路径及训练集标注信息）、val_dataloader（图片路径及验证集标注信息）、val_evaluator（验证集标注信息）、model（冻结骨干网络stages数，种类数）、param_scheduler（学习率衰减趋势）、optim_wrapper（学习率赋值）、default_hooks（模型权重保存策略）、custom_hooks（数据管道切换）、load_from（预训练权重加载路径）、train_cfg（赋值max_epochs以及验证测量）、randomness（固定随机数种子）、visualizer（选择可视化平台）

• 配置文件最重要的就是metainfo参数和model参数，一定要检查分类数是否正确，以及调色板数量是否一致。注意：即使只有1类，metainfo也要写成'classes': ('cat', ),括号中的逗号一定要有，否则报错。model中的bbox_head也要和种类数一致。

• 学习率缩放一般遵循经验法则：base_lr_default * (your_bs / default_bs)。从上面结构图中可以看到RTMDet模型有4个stages，model配置中dict(backbone=dict(frozen_stages=4), bbox_head=dict(num_classes=2))表示冻结了4个stages，即骨干网络全冻结。

config_animals = """
# Inherit and overwrite part of the config based on this config
_base_ = './rtmdet_l_8xb32-300e_coco.py'

data_root = './data/' # dataset root

train_batch_size_per_gpu = 24
train_num_workers = 4

max_epochs = 50
stage2_num_epochs = 6
base_lr = 0.000375


metainfo = {
    'classes': ('cat', 'dog', ),
    'palette': [
        (252, 215, 99), (153, 197, 252), 
    ]
}

train_dataloader = dict(
    batch_size=train_batch_size_per_gpu,
    num_workers=train_num_workers,
    dataset=dict(
        data_root=data_root,
        metainfo=metainfo,
        data_prefix=dict(img='images/'),
        ann_file='annotations/trainval.json'))

val_dataloader = dict(
    batch_size=train_batch_size_per_gpu,
    num_workers=train_num_workers,
    dataset=dict(
        data_root=data_root,
        metainfo=metainfo,
        data_prefix=dict(img='images/'),
        ann_file='annotations/trainval.json'))

test_dataloader = val_dataloader

val_evaluator = dict(ann_file=data_root + 'annotations/trainval.json')

test_evaluator = val_evaluator

model = dict(backbone=dict(frozen_stages=4), bbox_head=dict(num_classes=2))

# learning rate
param_scheduler = [
    dict(
        type='LinearLR',
        start_factor=1.0e-5,
        by_epoch=False,
        begin=0,
        end=1000),
    dict(
        # use cosine lr from 10 to 20 epoch
        type='CosineAnnealingLR',
        eta_min=base_lr * 0.05,
        begin=max_epochs // 2,
        end=max_epochs,
        T_max=max_epochs // 2,
        by_epoch=True,
        convert_to_iter_based=True),
]

train_pipeline_stage2 = [
    dict(type='LoadImageFromFile', backend_args=None),
    dict(type='LoadAnnotations', with_bbox=True),
    dict(
        type='RandomResize',
        scale=(640, 640),
        ratio_range=(0.1, 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='PackDetInputs')
]

# optimizer
optim_wrapper = dict(
    _delete_=True,
    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))

default_hooks = dict(
    checkpoint=dict(
        interval=5,
        max_keep_ckpts=2,  # only keep latest 2 checkpoints
        save_best='auto'
    ),
    logger=dict(type='LoggerHook', interval=20))

custom_hooks = [
    dict(
        type='PipelineSwitchHook',
        switch_epoch=max_epochs - stage2_num_epochs,
        switch_pipeline=train_pipeline_stage2)
]

# load COCO pre-trained weight
load_from = './checkpoints/rtmdet_l_8xb32-300e_coco_20220719_112030-5a0be7c4.pth'

train_cfg = dict(type='EpochBasedTrainLoop', max_epochs=max_epochs, val_begin=20, val_interval=1)
randomness = dict(seed=2023, deterministic=True, diff_rank_seed=False)
visualizer = dict(vis_backends=[dict(type='LocalVisBackend'), dict(type='WandbVisBackend')])
"""

with open('./configs/rtmdet/rtmdet_l_1xb4-100e_animals.py', 'w') as f:
    f.write(config_animals)

模型训练

• 做好上面的工作以后就可以开始模型训练了

!python tools/train.py configs/rtmdet/rtmdet_l_1xb4-100e_animals.py

• 模型epoch = 50时的精度

 Average Precision  (AP) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.952
 Average Precision  (AP) @[ IoU=0.50      | area=   all | maxDets=100 ] = 1.000
 Average Precision  (AP) @[ IoU=0.75      | area=   all | maxDets=100 ] = 0.995
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.800
 Average Precision  (AP) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.919
 Average Precision  (AP) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.959
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=  1 ] = 0.964
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets= 10 ] = 0.965
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=   all | maxDets=100 ] = 0.965
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= small | maxDets=100 ] = 0.800
 Average Recall     (AR) @[ IoU=0.50:0.95 | area=medium | maxDets=100 ] = 0.939
 Average Recall     (AR) @[ IoU=0.50:0.95 | area= large | maxDets=100 ] = 0.970
07/10 07:35:26 - mmengine - INFO - bbox_mAP_copypaste: 0.952 1.000 0.995 0.800 0.919 0.959
07/10 07:35:27 - mmengine - INFO - Epoch(val) [50][123/123]    coco/bbox_mAP: 0.9520  coco/bbox_mAP_50: 1.0000  coco/bbox_mAP_75: 0.9950  coco/bbox_mAP_s: 0.8000  coco/bbox_mAP_m: 0.9190  coco/bbox_mAP_l: 0.9590  data_time: 0.0532  time: 0.8068

• 我们可以打开wandb平台，跟踪训练精度，并将各项指标进行可视化
  
  

模型推理

• 当我们微调好模型后，可以在图片上进行推理

from mmdet.apis import DetInferencer
import glob

config = 'configs/rtmdet/rtmdet_l_1xb4-100e_animals.py'
checkpoint = glob.glob('./work_dirs/rtmdet_l_1xb4-100e_animals/best_coco*.pth')[0]

device = 'cuda:0'

inferencer = DetInferencer(config, checkpoint, device)

img = './data/images/Cats_Test1011.png'
result = inferencer(img, out_dir='./output', pred_score_thr=0.6)

display.clear_output()
Image.open('./output/vis/Cats_Test1011.png')

img = './data/images/Cats_Test1035.png'
result = inferencer(img, out_dir='./output', pred_score_thr=0.6)

display.clear_output()
Image.open('./output/vis/Cats_Test1035.png')