Detectron2学习笔记

简介

model_zoo/model_zoo.py内定义了很多预训练模型和对应的超参数

目标检测COCO-detection：Faster R-CNN, RetinaNet, RPN and Fast R-CNN

实例分割COCO-InstanceSegmentation: Mask R-CNN

姿态检测COCO-Person-Keypoint: keypoint R-CNN

全景分割COCO-PanopticSegmentation: Panoptic FPN

python demo/demo.py --config-file configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x.yaml \
 --input input1.jpg \
 --output . \
 --opts MODEL.WEIGHTS detectron2://COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_3x/137849600/model_final_f10217.pkl 

说明一下：
-config-file configs/COCO-… 读取 model 配置

–input input1.jpg 设置输入。 input1.jpg 位于根目录
–output . |设置输出路径，这里设为根目录

Detectron2库安装

win10安装detectron2_ymy_ymy22的博客-CSDN博客_detectron2 windowsWindows10下安装detectron2https://blog.csdn.net/ymy_ymy22/article/details/125285016?spm=1001.2014.3001.5501测试 demo/demo.py

Detectron2使用指南

深度学习 | Detectron2使用指南_JUST LOVE SMILE的博客-CSDN博客_detectron2使用文章目录1. Detectron2安装1.1 Linux1.2 Windows1.2.1 VS2019 C++编译环境1.2.2 pycocotools1.2.3 Detectron22. 自定义数据集2.1 关于COCO格式2.2 注册数据集2.3 可视化工具2.4 自定义数据增强3. 自定义模型3.1 特征提取网络（backbone）3.2 候选框生成器（proposal_generator）3.3 检测器（roi_heads）3.4 模型框架（meta_arch）4. 模型训练4.1 默认训练4.2https://blog.csdn.net/qq_43701912/article/details/123590274?utm_medium=distribute.pc_relevant.none-task-blog-2~default~baidujs_baidulandingword~default-0-123590274-blog-116749084.pc_relevant_default&spm=1001.2101.3001.4242.1&utm_relevant_index=2Detectron2的模型是分模块的，它将目标检测模型拆分为了4个核心模块：backbone，proposal_generator，roi_heads以及meta_arch

在detectron2/modeling/backbone 里有FPN，RegNet 和ResNet

detectron2/modeling里还有

 在projects里还有一些其他的任务网络

detectron2/project

 训练

tools/内有lightning_train_net, plain_train_net, train_net

tools/train_net.py --config-file configs/COCO-InstanceSegmentation/mask_rcnn_R_50_FPN_1x.yaml --num-gpus 1 SOLVER.IMS_PER_BATCH 2 SOLVER.BASE_LR 0.0025

制作自己的Detectron2预训练模型 - 知乎项目初期样本量较少，随着项目进展，样本数量不断增多，同一个项目需要不断迭代，训练多次，不用每次训练从0开始，可以基于前面的模型进行微调，缩短训练时间。 方法如下： import torch # 默认加载 net = torch.l…https://zhuanlan.zhihu.com/p/147336249detectron2 下载预训练模型中断导致无法运行_sysu_first_yasuo的博客-CSDN博客在使用detectron2的R101的faster-rcnn和retinanet时如果会自动加载预训练的R101 backbone解决方法有两个，一个是覆盖yaml中model weight的设定，设为空字符串：cfg.MODEL.WEIGHTS = ''另一个是用vpn把这个backbone下载下来再弄到本地，同样要修改cfg.MODEL.WEIGHTS为你下载下来的文件的地址...https://blog.csdn.net/weixin_44326452/article/details/106070988?utm_medium=distribute.pc_aggpage_search_result.none-task-blog-2~aggregatepage~first_rank_ecpm_v1~rank_v31_ecpm-3-106070988-null-null.pc_agg_new_rank&utm_term=detectron2%E6%80%8E%E6%A0%B7%E9%80%89%E6%8B%A9%E9%A2%84%E8%AE%AD%E7%BB%83%E6%A8%A1%E5%9E%8B&spm=1000.2123.3001.4430Detectron2 “快速开始” Detection Tutorial Colab Notebook 详细解读_Yukaz123的博客-CSDN博客Detectron快速上手官方 Colab Notebook 上 Getting Started 部分阅读：1. 使用预训练的Detectron2模型下载一张图片，我们需要创建detectron2 config，随后根据config创建一个Default Predictor去进行单张图片推理。cfg = get_cfg() # 获取 Default Config# 根据 mask_rcnn_R_50_FPN_3x.yaml 的配置文件更新 configcfg.merge_from_file(mohttps://blog.csdn.net/weixin_42174674/article/details/116289893

 常用命令

tools/train_net.py --config-file configs/COCO-InstanceSegmentation/mask_rcnn_R_50_C4_1x_VCM.yaml --num-gpus 1 

python demo\demo.py --input input.jpg

python tools/analyze_model.py --task flop --config-file configs/COCO-Detection/faster_rcnn_R_50_C4_1x.yaml

mmdet转detectron2

# Copyright (c) OpenMMLab. All rights reserved.
import argparse
from collections import OrderedDict

import torch
import argparse
from collections import OrderedDict

import mmcv
import torch

arch_settings = {50: (3, 4, 6, 3), 101: (3, 4, 23, 3)}


def convert_bn(blobs, state_dict, caffe_name, torch_name, converted_names):
    # detectron replace bn with affine channel layer
    state_dict[torch_name + '.bias'] = blobs[caffe_name +'.bias']
    state_dict[torch_name + '.weight'] = blobs[caffe_name +'.weight']
    # bn_size = state_dict[torch_name + '.weight'].size()
    state_dict[torch_name + '.running_mean'] = blobs[caffe_name +'.running_mean']
    state_dict[torch_name + '.running_var'] = blobs[caffe_name + '.running_var']
    converted_names.add(caffe_name + '.bias')
    converted_names.add(caffe_name + '.weight')
    converted_names.add(caffe_name +'.running_mean')
    converted_names.add(caffe_name + '.running_var')


def convert_conv_fc(blobs, state_dict, caffe_name, torch_name, converted_names):
    state_dict[torch_name + '.weight'] = blobs[caffe_name +'.weight']
    converted_names.add(caffe_name + '.weight')
    if caffe_name + '.bias' in blobs:
        state_dict[torch_name + '.bias'] = blobs[caffe_name +'.bias']
        converted_names.add(caffe_name + '.bias')


def convert(src, dst, depth):
    """Convert keys in detectron pretrained ResNet models to pytorch style."""
    # load arch_settings
    if depth not in arch_settings:
        raise ValueError('Only support ResNet-50 and ResNet-101 currently')
    block_nums = arch_settings[depth]
    # load caffe model
    caffe_model = torch.load(src, map_location='cpu') #mmcv.load(src, encoding='latin1')
    blobs = caffe_model['blobs'] if 'blobs' in caffe_model else caffe_model
    blobs = blobs['state_dict']
    # convert to pytorch style
    state_dict = OrderedDict()
    converted_names = set()

    convert_conv_fc(blobs, state_dict, 'roi_head.bbox_head.fc_cls', 'roi_heads.box_predictor.cls_score', converted_names)
    convert_conv_fc(blobs, state_dict, 'roi_head.bbox_head.fc_reg', 'roi_heads.box_predictor.bbox_pred', converted_names)
    convert_conv_fc(blobs, state_dict, 'roi_head.bbox_head.shared_fcs.0', 'roi_heads.box_head.fc1', converted_names)
    convert_conv_fc(blobs, state_dict, 'roi_head.bbox_head.shared_fcs.1', 'roi_heads.box_head.fc2', converted_names)

    convert_conv_fc(blobs, state_dict, 'rpn_head.rpn_conv', 'proposal_generator.rpn_head.conv', converted_names)
    convert_conv_fc(blobs, state_dict, 'rpn_head.rpn_cls', 'proposal_generator.rpn_head.objectness_logits', converted_names)
    convert_conv_fc(blobs, state_dict, 'rpn_head.rpn_reg', 'proposal_generator.rpn_head.anchor_deltas', converted_names)

    convert_conv_fc(blobs, state_dict, 'backbone.conv1', 'backbone.bottom_up.stem.conv1', converted_names)
    convert_bn(blobs, state_dict, 'backbone.bn1', 'backbone.bottom_up.stem.conv1.norm', converted_names)

    for i in range(0, 4):
        convert_conv_fc(blobs, state_dict, f'neck.lateral_convs.{i}.conv', f'backbone.fpn_lateral{i+2}', converted_names)
        convert_conv_fc(blobs, state_dict, f'neck.fpn_convs.{i}.conv', f'backbone.fpn_output{i + 2}', converted_names)


    for i in range(1, len(block_nums) + 1):
        for j in range(block_nums[i - 1]):
            if j == 0:
                convert_conv_fc(blobs, state_dict, f'backbone.layer{i}.{j}.downsample.0', f'backbone.bottom_up.res{i + 1}.{j}.shortcut', converted_names)
                convert_bn(blobs, state_dict, f'backbone.layer{i}.{j}.downsample.1', f'backbone.bottom_up.res{i + 1}.{j}.shortcut.norm', converted_names)
            for k in range(1, 4):
                convert_conv_fc(blobs, state_dict, f'backbone.layer{i}.{j}.conv{k}', f'backbone.bottom_up.res{i + 1}.{j}.conv{k}',converted_names)
                convert_bn(blobs, state_dict, f'backbone.layer{i}.{j}.bn{k}', f'backbone.bottom_up.res{i + 1}.{j}.conv{k}.norm',  converted_names)
    # check if all layers are converted
    for key in blobs:
        if key not in converted_names:
            print(f'Not Convert: {key}')
    # save checkpoint
    checkpoint = dict()
    checkpoint['model'] = state_dict
    torch.save(checkpoint, dst)



def main():
    parser = argparse.ArgumentParser(description='Convert model keys')
    parser.add_argument('src', help='src detectron model path')
    parser.add_argument('dst', help='save path')
    parser.add_argument('depth', type=int, help='ResNet model depth')
    args = parser.parse_args()
    convert(src, dst, depth)



if __name__ == '__main__':
    main()

注意：coco数据集的PIXEL_STD需要修改