mmdetection 使用maskrcnn训练自己的coco数据集实现目标检测【过程记录】更新中

mmdetection 使用maskrcnn训练自己的coco数据集实现目标检测【过程记录】

ZZZZ_Y_

2075人浏览 · 2023-10-20 23:45:39

ZZZZ_Y_ · 2023-10-20 23:45:39 发布

搭建环境

搭建mmdetection环境：mmdetection环境搭建（linux）-CSDN博客

数据集

将数据集放在detection-master/data/coco下

配置文件config设置

看到网上有很多方法是直接修改的官方的模型院配置文件，但是我不太想修改原文件，一是麻烦，得改很多文件，二是担心越改越乱容易出问题，于是选择配置自己的config文件，训练时直接调用自己的config文件，不修改官方提供的模型原配置文件。

目录 mmdetection-master/configs/ 下有各模型文件夹，文件夹下为各模型的配置文件，如下图所示这里我使用mask_rcnn_r50_fpn_1x_coco.py实现目标检测。

准备工作

在 mmdetection-master/configs/ 下新建demo文件夹，用于存放自己的config文件，这里将自己的config文件命名为demo_mask_rcnn_r50_fpn_1x_coco.py

在mmdetection-master下新建checkpoints文件夹，将要用的预训练模型权重文件下载到此文件夹下

开始配置自己的config文件

首先继承基础的模型结构

_base_ = '../mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py'

然后修改配置文件中的各项参数完成继承

由于要实现分类任务，因此需要对模型结构进行修改：

1. 删掉mask_head和mask_roi_extractor，通过将对应节点设置为key即可删除
2. 修改bbox_head的输出，即输出类别数，这里我的类别数是5

model = dict(
    roi_head=dict(
        bbox_head=dict(num_classes=5),
        mask_head=None,
        mask_roi_extractor=None))

采用的数据集为coco数据集

dataset_type = 'COCODataset'

设置检测类别

classes = ('Inlet', 'Slightshort', 'Generalshort', 'Severeshort', 'Outlet')

修改配置文件中的中间变量

1.由于仅实现检测任务，将train_pipeline中的with_mask设置为False，默认为True
2.train_pipeline中的 dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels', 'gt_labels'])为默认设置，删掉'gt_labels'

ps. train中要加上pipeline = train_pipeline将train_pipeline传递到data里

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True, with_mask=False),
    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(
        type='Normalize',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        to_rgb=True),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]

设置数据集路径、batchsize、num_workers等参数，train中要加上pipeline = train_pipeline将train_pipeline传递到data里

data = dict(
    samples_per_gpu=8,  # batch size
    workers_per_gpu=4,  # num_workers
    train=dict(
        #img_prefix='/data/zy/dataset/project/Cooper001_withlabel/coco/train2017/',
        img_prefix='/data/zy/dataset/project/Cooper001_withlabel/coco/train2017_alpha2/',
        #img_prefix='/data/zy/dataset/project/Cooper001_withlabel/coco/train2017_rec/',
        classes=classes,
        pipeline = train_pipeline, #这里将前面修改的train_pipeline传递到pipeline里，不然修改不生效
        ann_file='/data/zy/dataset/project/Cooper001_withlabel/coco/annotations/instances_train2017.json'),
    val=dict(
        img_prefix='/data/zy/dataset/project/Cooper001_withlabel/coco/val2017_alpha2/',
        classes=classes,
        ann_file='/data/zy/dataset/project/Cooper001_withlabel/coco/annotations/instances_val2017.json'),
    test=dict(
        img_prefix='/data/zy/dataset/project/Cooper001_withlabel/coco/test2017_alpha2/',
        classes=classes,
        ann_file='/data/zy/dataset/project/Cooper001_withlabel/coco/annotations/instances_test2017.json'))

evalua的metric设为bbox，classwise设为True打印每个类别的ap

evaluation = dict(interval=1, metric='bbox', classwise=True)

设置epoch

runner = dict(type='EpochBasedRunner', max_epochs=20)

optimizer

optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)

加载预训练模型权重文件

load_from = '/data/zy/code/mmdetection-master/checkpoints/mask_rcnn_r50_fpn_1x_coco_20200205-d4b0c5d6.pth'

最终的config文件

_base_ = '../mask_rcnn/mask_rcnn_r50_fpn_1x_coco.py'


model = dict(
    roi_head=dict(
        bbox_head=dict(num_classes=5),
        mask_head=None,
        mask_roi_extractor=None))
        
      
evaluation = dict(interval=1, metric='bbox', classwise=True)  

#optimizer = dict(type='SGD', lr=0.02, momentum=0.9, weight_decay=0.0001)

train_pipeline = [
    dict(type='LoadImageFromFile'),
    dict(type='LoadAnnotations', with_bbox=True, with_mask=False),
    dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
    dict(type='RandomFlip', flip_ratio=0.5),
    dict(
        type='Normalize',
        mean=[123.675, 116.28, 103.53],
        std=[58.395, 57.12, 57.375],
        to_rgb=True),
    dict(type='Pad', size_divisor=32),
    dict(type='DefaultFormatBundle'),
    dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels'])
]

runner = dict(type='EpochBasedRunner', max_epochs=20)

dataset_type = 'COCODataset'
classes = ('Inlet', 'Slightshort', 'Generalshort', 'Severeshort', 'Outlet')

data = dict(
    samples_per_gpu=8,  # batch size
    workers_per_gpu=4,  # num_workers
    train=dict(
        img_prefix='/data/zy/dataset/project/Cooper001_withlabel/coco/train2017/',        
        classes=classes,
        pipeline = train_pipeline,
        ann_file='/data/zy/dataset/project/Cooper001_withlabel/coco/annotations/instances_train2017.json'),
    val=dict(
        img_prefix='/data/zy/dataset/project/Cooper001_withlabel/coco/val2017/',
        classes=classes,
        ann_file='/data/zy/dataset/project/Cooper001_withlabel/coco/annotations/instances_val2017.json'),
    test=dict(
        img_prefix='/data/zy/dataset/project/Cooper001_withlabel/coco/test2017/',
        classes=classes,
        ann_file='/data/zy/dataset/project/Cooper001_withlabel/coco/annotations/instances_test2017.json'))
 


load_from = '/data/zy/code/mmdetection-master/checkpoints/mask_rcnn_r50_fpn_1x_coco_20200205-d4b0c5d6.pth'

训练

在mmdetection-master下新建文件夹work_dirs，用于存放训练过程生成的文件

修改tools/train.py文件中的config和work-dir设置

    parser.add_argument('--config', default='./configs/demo/demo_faster_rcnn_r50_fpn_2x_coco.py', help='train config file path')
    parser.add_argument('--work-dir', default='./work_dirs/', help='the dir to save logs and models')

命令行输入一下命令进行训练，--config后为自己的配置文件路径，--work-dir后为用于存放训练过程生成的文件的文件夹路径

python tools/train.py --config ./configs/demo/demo_mask_rcnn_r50_fpn_1x_coco.py --work-dir ./work_dirs/maskrcnn

开始训练

绘制map曲线

python tools/analysis_tools/analyze_logs.py plot_curve ./work_dirs/maskrcnn_orig/20231030_153716.log.json --keys bbox_mAP_50 --out ./work_dirs/maskrcnn_orig/map.jpg

测试

python tools/test.py \
--config configs/demo/demo_mask_rcnn_r50_fpn_1x_coco.py \
--work-dir work_dirs/maskrcnn/ \
--checkpoint  work_dirs/maskrcnn/latest.pth \
--eval bbox \
--out work_dirs/maskrcnn/result.pkl \
--show-score-thr 0.5 \  #这一行应该可以删掉
--show-dir work_dirs/maskrcnn/infered_images/ \ # 保存推理后的图片
--eval-options "classwise=True" # 计算类别的AP

Visualize

修改标签文字背景颜色、字体颜色、位置等

在mmdetection-master/mmdet/core/visualization/image.py中找到draw_labels函数修改facecolor

def draw_labels(ax,
                labels,
                positions,
                scores=None,
                class_names=None,
                color='w',
                font_size=12,
                scales=None,
                horizontal_alignment='left'):
    """Draw labels on the axes.

    Args:
        ax (matplotlib.Axes): The input axes.
        labels (ndarray): The labels with the shape of (n, ).
        positions (ndarray): The positions to draw each labels.
        scores (ndarray): The scores for each labels.
        class_names (list[str]): The class names.
        color (list[tuple] | matplotlib.color): The colors for labels.
        font_size (int): Font size of texts. Default: 8.
        scales (list[float]): Scales of texts. Default: None.
        horizontal_alignment (str): The horizontal alignment method of
            texts. Default: 'left'.

    Returns:
        matplotlib.Axes: The result axes.
    """
    for i, (pos, label) in enumerate(zip(positions, labels)):
        label_text = class_names[
            label] if class_names is not None else f'class {label}'
        if scores is not None:
            label_text += f'|{scores[i]:.02f}'
        text_color = color[i] if isinstance(color, list) else color

        font_size_mask = font_size if scales is None else font_size * scales[i]
        ax.text(
            pos[0],
            pos[1],
            f'{label_text}',
            bbox={
                'facecolor': 'white',    # here to rectify
                'alpha': 0.8,
                'pad': 0.7,
                'edgecolor': 'none'
            },
            color=text_color,
            fontsize=font_size_mask,
            verticalalignment='top',
            horizontalalignment=horizontal_alignment)

    return ax

修改bbox相关

def draw_bboxes(ax, bboxes, color='g', alpha=0.8, thickness=2):
    """Draw bounding boxes on the axes.

    Args:
        ax (matplotlib.Axes): The input axes.
        bboxes (ndarray): The input bounding boxes with the shape
            of (n, 4).
        color (list[tuple] | matplotlib.color): the colors for each
            bounding boxes.
        alpha (float): Transparency of bounding boxes. Default: 0.8.
        thickness (int): Thickness of lines. Default: 2.

    Returns:
        matplotlib.Axes: The result axes.
    """
    polygons = []
    for i, bbox in enumerate(bboxes):
        bbox_int = bbox.astype(np.int32)
        poly = [[bbox_int[0], bbox_int[1]], [bbox_int[0], bbox_int[3]],
                [bbox_int[2], bbox_int[3]], [bbox_int[2], bbox_int[1]]]
        np_poly = np.array(poly).reshape((4, 2))
        polygons.append(Polygon(np_poly))
    p = PatchCollection(
        polygons,
        facecolor='none',
        edgecolors=color,
        linewidths=thickness,
        alpha=alpha)
    ax.add_collection(p)

    return ax

下面这些文件里面可能也有可修改的地方，但是没有研究

mmdetection-master/mmdet/apis/inference.py

mmdetection-master/mmdet/apis/test.py

参考文件：

官方教程教程 1: 学习配置文件 — MMDetection 2.21.0 文档