网站推广的方法和手段,学做网站论坛会员账号,无货源电商怎么起步,dedecms网站入侵前言
MMPose是一款基于PyTorch的姿态分析开源工具箱#xff0c;是OpenMMLab项目成员之一#xff0c;主要特性#xff1a; 支持多种人体姿态分析相关任务#xff1a;2D多人姿态估计、2D手部姿态估计、动物关键点检测等等更高的精度和更快的速度#xff1a;包括“自顶向下”…前言
MMPose是一款基于PyTorch的姿态分析开源工具箱是OpenMMLab项目成员之一主要特性 支持多种人体姿态分析相关任务2D多人姿态估计、2D手部姿态估计、动物关键点检测等等更高的精度和更快的速度包括“自顶向下”和“自底向上”两大类算法支持多样的数据集支持了很多主流数据集的准备和构建如 COCO、 MPII等模块化设计将统一的人体姿态分析框架解耦成不同的模块组件通过组合不同的模块组件可以便捷地构建自定义人体姿态分析模型 本文主要对动物关键点检测模型进行微调与测试从数据集构造开始详细解释各模块作用。对一些新手可能会犯的错误做一些说明本文使用的数据集为kaggle平台中Cat Dataset数据集数据说明环境为kaggle平台提供的P100 GPU完整的Jupyter Notebook放在这里欢迎大家Copy Edit
环境配置
mmcv的安装方式在我前面的mmdetection和mmsegmentation教程中都有写到。这里不再提MMPose安装方法最好是使用git如果没有git工具可以使用mim install mmpose 最后在项目文件夹下新建checkpoint、outputs、data文件夹分别用来存放模型预训练权重、模型输出结果、训练数据
from IPython import display
!pip install openmim
!pip install -q /kaggle/input/frozen-packages-mmdetection/mmcv-2.0.1-cp310-cp310-linux_x86_64.whl!git clone https://github.com/open-mmlab/mmdetection.git
%cd mmdetection
!pip install -e .%cd ..
!git clone https://github.com/open-mmlab/mmpose.git
%cd mmpose
!pip install -e .!mkdir checkpoint
!mkdir outputs
!mkdir data
display.clear_output()在上面的安装工作完成后我们检查一下环境以及核对一下安装版本
from IPython import display
import mmcv
from mmcv.ops import get_compiling_cuda_version, get_compiler_version
print(MMCV版本, mmcv.__version__)
%cd /kaggle/working/mmdetection
import mmdet
print(mmdetection版本, mmdet.__version__)
%cd /kaggle/working/mmpose
import mmpose
print(mmpose版本, mmpose.__version__)
print(CUDA版本, get_compiling_cuda_version())
print(编译器版本, get_compiler_version())输出
MMCV版本 2.0.1
/kaggle/working/mmdetection
mmdetection版本 3.1.0
/kaggle/working/mmpose
mmpose版本 1.1.0
CUDA版本 11.8
编译器版本 GCC 11.3为方便后续进行文件操作导入一些常用库
import os
import io
import json
import shutil
import random
import numpy as np
from pathlib import Pathfrom PIL import Image
from tqdm import tqdm
from mmengine import Configfrom pycocotools.coco import COCO预训练模型推理
在进行姿态估计前需要目标检测将不同的目标检测出来然后再分别对不同的目标进行姿态估计。所以我们要选择一个目标检测模型。这里选择的是mmdetection工具箱中的RTMDet模型型号选择RTMDet-l。配置文件位于mmdetection/configs/rtmdet/rtmdet_l_8xb32-300e_coco.py我们复制模型权重地址并进行下载。 姿态估计模型选择RTMPose模型打开mmpose项目文件夹projects/rtmpose/README.md文档发现RTMPose模型动物姿态估计Animal 2d (17 Keypoints)仅提供了一个预训练模型。配置文件位于projects/rtmpose/rtmpose/animal_2d_keypoint/rtmpose-m_8xb64-210e_ap10k-256x256.py我们复制模型权重地址并进行下载。 将预训练权重模型全部放入mmpose项目文件夹的checkpoint文件夹下。
# 下载RTMDet-L模型用于目标检测
!wget https://download.openmmlab.com/mmdetection/v3.0/rtmdet/rtmdet_l_8xb32-300e_coco/rtmdet_l_8xb32-300e_coco_20220719_112030-5a0be7c4.pth -P checkpoint
# 下载RTMPose模型用于姿态估计
!wget https://download.openmmlab.com/mmpose/v1/projects/rtmposev1/rtmpose-m_simcc-ap10k_pt-aic-coco_210e-256x256-7a041aa1_20230206.pth -P checkpoint
display.clear_output()MMPose提供了一个被称为MMPoseInferencer的、全面的推理API。这个API使得用户得以使用所有MMPose支持的模型来对图像和视频进行模型推理。此外该API可以完成推理结果自动化并方便用户保存预测结果。我们使用Cat Dataset数据集中的一张图片作为示例进行模型推理。推理参数说明 det_modelmmdetection工具箱中目标检测模型配置文件det_weightsmmdetection工具箱中目标检测模型对应预训练权重文件pose2dmmpose工具箱中姿态估计模型配置文件pose2d_weightsmmpose工具箱中姿态估计对应预训练权重文件out_dir图片生成的文件夹
from mmpose.apis import MMPoseInferencerimg_path /kaggle/input/cat-dataset/CAT_00/00000001_012.jpg
# 使用模型别名创建推断器
inferencer MMPoseInferencer(det_model /kaggle/working/mmdetection/configs/rtmdet/rtmdet_l_8xb32-300e_coco.py,det_weights checkpoint/rtmdet_l_8xb32-300e_coco_20220719_112030-5a0be7c4.pth,pose2d projects/rtmpose/rtmpose/animal_2d_keypoint/rtmpose-m_8xb64-210e_ap10k-256x256.py,pose2d_weights checkpoint/rtmpose-m_simcc-ap10k_pt-aic-coco_210e-256x256-7a041aa1_20230206.pth)# MMPoseInferencer采用了惰性推断方法在给定输入时创建一个预测生成器
result_generator inferencer(img_path, out_dir outputs, showFalse)
result next(result_generator)
display.clear_output()可视化推理结果
import matplotlib.pyplot as pltimg_og mmcv.imread(img_path)
img_fuse mmcv.imread(outputs/visualizations/00000001_012.jpg)fig, axes plt.subplots(1, 2, figsize(15, 10))
axes[0].imshow(mmcv.bgr2rgb(img_og))
axes[0].set_title(Original Image)
axes[0].axis(off)axes[1].imshow(mmcv.bgr2rgb(img_fuse))
axes[1].set_title(Keypoint Image)
axes[1].axis(off)
plt.show()数据处理
数据内容详解
Cat Dataset包含9000多张猫图像。对于每张图像都有猫头部的注释有9个点2个用于眼睛1个用于嘴巴6个用于耳朵。注释数据存储在1个文件中文件名是相应的图像名称末尾加上“cat”。每张猫图像都有1个注释文件。对于每个注释文件注释数据按以下顺序存储 Number of points 关键点数目Left Eye左眼Right Eye右眼Mouth嘴Left Ear-1左耳-1Left Ear-2左耳-2Left Ear-3左耳-3Right Ear-1右耳-1Right Ear-2右耳-2Right Ear-3左耳-3 数据集最初在互联网档案馆中找到网站数据层级目录如下所示 - CAT_00- 00000001_000.jpg- 00000001_000.jpg.cat- 00000001_005.jpg- 00000001_005.jpg.cat- ...- CAT_01- 00000100_002.jpg- 00000100_002.jpg.cat- 00000100_003.jpg- 00000100_003.jpg.cat- CAT_02- CAT_03- CAT_04- CAT_05- CAT_06总的来说一共有7个文件夹每个文件夹里面有若干.jpg格式的图片文件且对应有.cat格式的注释文件.cat文件可以看做是文本文件内容示例
9 435 322 593 315 524 446 318 285 283 118 430 195 568 186 701 81 703 267 除第1个数字9表示有9个关键点后面每2个点表示1个部位的坐标(x,y)所以一共有1 2 * 9 19个点
文件夹规整
我们将数据集中的7个文件夹中的图片与注释文件分开分别存储在mmpose项目文件夹data文件夹中并分别命名为images、ann
def separate_files(og_folder, trans_folder):image_folder os.path.join(trans_folder, images)ann_folder os.path.join(trans_folder, ann)os.makedirs(image_folder, exist_okTrue)os.makedirs(ann_folder, exist_okTrue)for folder in os.listdir(data_folder):folder_path os.path.join(data_folder, folder)if os.path.isdir(folder_path):for file in os.listdir(folder_path):if file.endswith(.jpg):source_path os.path.join(folder_path, file)target_path os.path.join(image_folder, file)shutil.copy(source_path, target_path)elif file.endswith(.cat):source_path os.path.join(folder_path, file)target_path os.path.join(ann_folder, file)shutil.copy(source_path, target_path)data_folder /kaggle/input/cat-dataset
trans_folder ./dataseparate_files(data_folder, trans_folder)构造COCO注释文件
本质上来说COCO就是1个字典文件第1级键包含images、annotations、categories。 其中images包含id图片的唯一标识必须要是数值型不能有字符 、file_name图片名字、 height图片高度, width图片宽度这些信息其中annotations包含category_id图片所属种类、segmentation实例分割掩码、iscrowd决定是RLE格式还是polygon格式、image_id图片id对应images键中的id、id注释信息id、bbox目标检测框[x, y, width, height]、 area目标检测框面积、num_keypoints关键点数量, keypoints关键点坐标其中categories包含supercategory、id类别id、name类别名、keypoints各部位名称、skeleton部位连接信息 更详细的COCO注释文件解析推荐博客COCO数据集的标注格式、如何将VOC XML文件转化成COCO数据格式构造read_file_as_list函数将注释文件中的坐标变成[x,y,v]v为0时表示这个关键点没有标注v为1时表示这个关键点标注了但是不可见被遮挡了v为2时表示这个关键点标注了同时可见。因为数据集中部位坐标均标注且可见所以在x,y坐标后均插入2。
def read_file_as_list(file_path):with open(file_path, r) as file:content file.read()key_point [int(num) for num in content.split()]key_num key_point[0]key_point.pop(0)for i in range(2, len(key_point) len(key_point)//2, 2 1):key_point.insert(i, 2)return key_num,key_point构造get_image_size函数用于获取图片宽度和高度。
def get_image_size(image_path):with Image.open(image_path) as img:width, height img.sizereturn width, height因为数据集没有提供目标检测框信息且图片中基本无干扰元素所以将目标检测框信息置为[0, 0, width, height]即整张图片。相应的目标检测框面积area width * height。
# 转换为coco数据格式
def coco_structure(ann_dir,image_dir):coco dict()coco[images] []coco[annotations] []coco[categories] []coco[categories].append(dict(supercategory cat,id 1,name cat,keypoints [Left Eye,Right Eye,Mouth,Left Ear-1,Left Ear-2,Left Ear-3,Right Ear-1,Right Ear-2,Right Ear-3],skeleton [[0,1],[0,2],[1,2],[3,4],[4,5],[5,6],[6,7],[7,8],[3,8]]))ann_list os.listdir(ann_dir)id 0for file_name in tqdm(ann_list):key_num,key_point read_file_as_list(os.path.join(ann_dir, file_name))if key_num 9:image_name os.path.splitext(file_name)[0]image_id os.path.splitext(image_name)[0]height, width get_image_size(os.path.join(image_dir, image_name))image {id: id, file_name: image_name, height: height, width: width}coco[images].append(image)key_dict dict(category_id 1, segmentation [], iscrowd 0, image_id id, id id, bbox [0, 0, width, height], area width * height, num_keypoints key_num, keypoints key_point)coco[annotations].append(key_dict)id id 1return coco写入注释信息并将其保存为mmpose项目文件夹data/annotations_all.json文件
ann_file coco_structure(./data/ann,./data/images)
output_file_path ./data/annotations_all.json
with open(output_file_path, w, encodingutf-8) as output_file:json.dump(ann_file, output_file, ensure_asciiTrue, indent4)拆分训练、测试数据
按0.85、0.15的比例将注释文件拆分为训练、测试文件
def split_coco_dataset(coco_json_path: str, save_dir: str, ratios: list,shuffle: bool, seed: int):if not Path(coco_json_path).exists():raise FileNotFoundError(fCan not not found {coco_json_path})if not Path(save_dir).exists():Path(save_dir).mkdir(parentsTrue)ratios np.array(ratios) / np.array(ratios).sum()if len(ratios) 2:ratio_train, ratio_test ratiosratio_val 0train_type trainvalelif len(ratios) 3:ratio_train, ratio_val, ratio_test ratiostrain_type trainelse:raise ValueError(ratios must set 2 or 3 group!)coco COCO(coco_json_path)coco_image_ids coco.getImgIds()val_image_num int(len(coco_image_ids) * ratio_val)test_image_num int(len(coco_image_ids) * ratio_test)train_image_num len(coco_image_ids) - val_image_num - test_image_numprint(Split info: \nfTrain ratio {ratio_train}, number {train_image_num}\nfVal ratio {ratio_val}, number {val_image_num}\nfTest ratio {ratio_test}, number {test_image_num})seed int(seed)if seed ! -1:print(fSet the global seed: {seed})np.random.seed(seed)if shuffle:print(shuffle dataset.)random.shuffle(coco_image_ids)train_image_ids coco_image_ids[:train_image_num]if val_image_num ! 0:val_image_ids coco_image_ids[train_image_num:train_image_num val_image_num]else:val_image_ids Nonetest_image_ids coco_image_ids[train_image_num val_image_num:]categories coco.loadCats(coco.getCatIds())for img_id_list in [train_image_ids, val_image_ids, test_image_ids]:if img_id_list is None:continueimg_dict {images: coco.loadImgs(idsimg_id_list),categories: categories,annotations: coco.loadAnns(coco.getAnnIds(imgIdsimg_id_list))}if img_id_list train_image_ids:json_file_path Path(save_dir, f{train_type}.json)elif img_id_list val_image_ids:json_file_path Path(save_dir, val.json)elif img_id_list test_image_ids:json_file_path Path(save_dir, test.json)else:raise ValueError(img_id_list ERROR!)print(fSaving json to {json_file_path})with open(json_file_path, w) as f_json:json.dump(img_dict, f_json, ensure_asciiFalse, indent2)print(All done!)split_coco_dataset(./data/annotations_all.json, ./data, [0.85,0.15], True, 2023)输出
loading annotations into memory...
Done (t0.13s)
creating index...
index created!
Split info:
Train ratio 0.85, number 8495
Val ratio 0, number 0
Test ratio 0.15, number 1498
Set the global seed: 2023
shuffle dataset.
Saving json to data/trainval.json
Saving json to data/test.json
All done!可以看到训练集有8495张图片测试集有1498张图片
模型配置文件
打开项目文件夹下的projects/rtmpose/rtmpose/animal_2d_keypoint/rtmpose-m_8xb64-210e_ap10k-256x256.py文件发现模型配置文件仅继承_base_/default_runtime.py文件需要修改主要有dataset_type、data_mode、dataset_info、codec、train_dataloader 、test_dataloader 、val_evaluator、base_lr、max_epochs、default_hooks。还有一些细节我在代码中有标注可以参照着修改修改完成后将文件写入./configs/animal_2d_keypoint/cat_keypoint.py中
custom_config
_base_ [mmpose::_base_/default_runtime.py]# 数据集类型及路径
dataset_type CocoDataset
data_mode topdown
data_root ./data/
work_dir ./work_dir# cat dataset关键点检测数据集-元数据
dataset_info {dataset_name:Keypoint_cat,classes:cat,paper_info:{author:Luck,title:Cat Keypoints Detection,},keypoint_info:{0:{name:Left Eye,id:0,color:[255,0,0],type: ,swap: },1:{name:Right Eye,id:1,color:[255,127,0],type: ,swap: },2:{name:Mouth,id:2,color:[255,255,0],type: ,swap: },3:{name:Left Ear-1,id:3,color:[0,255,0],type: ,swap: },4:{name:Left Ear-2,id:4,color:[0,255,255],type: ,swap: },5:{name:Left Ear-3,id:5,color:[0,0,255],type: ,swap: },6:{name:Right Ear-1,id:6,color:[139,0,255],type: ,swap: },7:{name:Right Ear-2,id:7,color:[255,0,255],type: ,swap: },8:{name:Right Ear-3,id:8,color:[160,82,45],type: ,swap: }},skeleton_info: {0: {link:(Left Eye,Right Eye),id: 0,color: [255,0,0]},1: {link:(Left Eye,Mouth),id: 1,color: [255,0,0]},2: {link:(Right Eye,Mouth),id: 2,color: [255,0,0]},3: {link:(Left Ear-1,Left Ear-2),id: 3,color: [255,0,0]},4: {link:(Left Ear-2,Left Ear-3),id: 4,color: [255,0,0]},5: {link:(Left Ear-3,Right Ear-1),id: 5,color: [255,0,0]},6: {link:(Right Ear-1,Right Ear-2),id: 6,color: [255,0,0]},7: {link:(Right Ear-2,Right Ear-3),id: 7,color: [255,0,0]},8: {link:(Left Ear-1,Right Ear-3),id: 8,color: [255,0,0]},}
}# 获取关键点个数
NUM_KEYPOINTS len(dataset_info[keypoint_info])
dataset_info[joint_weights] [1.0] * NUM_KEYPOINTS
dataset_info[sigmas] [0.025] * NUM_KEYPOINTS# 训练超参数
max_epochs 100
val_interval 5
train_cfg {max_epochs: max_epochs, val_begin:20, val_interval: val_interval}
train_batch_size 32
val_batch_size 32
stage2_num_epochs 10
base_lr 4e-3 / 16
randomness dict(seed2023)# 优化器
optim_wrapper dict(typeOptimWrapper,optimizerdict(typeAdamW, lrbase_lr, weight_decay0.05),paramwise_cfgdict(norm_decay_mult0, bias_decay_mult0, bypass_duplicateTrue))# 学习率
param_scheduler [dict(typeLinearLR, start_factor1.0e-5, by_epochFalse, begin0, end600),dict(typeCosineAnnealingLR,eta_minbase_lr * 0.05,beginmax_epochs // 2,endmax_epochs,T_maxmax_epochs // 2,by_epochTrue,convert_to_iter_basedTrue),
]# automatically scaling LR based on the actual training batch size
auto_scale_lr dict(base_batch_size1024)# codec settings
# input_size可以换成128的倍数
# sigma高斯分布标准差越大越易学但进度低。高精度场景可以调小RTMPose 原始论文中为 5.66
# input_size、sigma和下面model中的in_featuremap_size参数需要成比例缩放
codec dict(typeSimCCLabel,input_size(512, 512),sigma(24, 24),simcc_split_ratio2.0,normalizeFalse,use_darkFalse)# 模型RTMPose-M
model dict(typeTopdownPoseEstimator,data_preprocessordict(typePoseDataPreprocessor,mean[123.675, 116.28, 103.53],std[58.395, 57.12, 57.375],bgr_to_rgbTrue),backbonedict(_scope_mmdet,typeCSPNeXt,archP5,expand_ratio0.5,deepen_factor0.67,widen_factor0.75,out_indices(4, ),channel_attentionTrue,norm_cfgdict(typeSyncBN),act_cfgdict(typeSiLU),init_cfgdict(typePretrained,prefixbackbone.,checkpointhttps://download.openmmlab.com/mmdetection/v3.0/rtmdet/cspnext_rsb_pretrain/cspnext-m_8xb256-rsb-a1-600e_in1k-ecb3bbd9.pth)),headdict(typeRTMCCHead,in_channels768,out_channelsNUM_KEYPOINTS,input_sizecodec[input_size],in_featuremap_size(16, 16),simcc_split_ratiocodec[simcc_split_ratio],final_layer_kernel_size7,gau_cfgdict(hidden_dims256,s128,expansion_factor2,dropout_rate0.,drop_path0.,act_fnSiLU,use_rel_biasFalse,pos_encFalse),lossdict(typeKLDiscretLoss,use_target_weightTrue,beta10.,label_softmaxTrue),decodercodec),test_cfgdict(flip_testTrue))backend_args dict(backendlocal)# pipelines
train_pipeline [dict(typeLoadImage, backend_argsbackend_args),dict(typeGetBBoxCenterScale),dict(typeRandomFlip, directionhorizontal),# dict(typeRandomHalfBody),dict(typeRandomBBoxTransform, scale_factor[0.8, 1.2], rotate_factor30),dict(typeTopdownAffine, input_sizecodec[input_size]),dict(typemmdet.YOLOXHSVRandomAug),dict(typeAlbumentation,transforms[dict(typeChannelShuffle, p0.5),dict(typeCLAHE, p0.5),# dict(typeDownscale, scale_min0.7, scale_max0.9, p0.2),dict(typeColorJitter, p0.5),dict(typeCoarseDropout,max_holes4,max_height0.3,max_width0.3,min_holes1,min_height0.2,min_width0.2,p0.5),]),dict(typeGenerateTarget, encodercodec),dict(typePackPoseInputs)
]val_pipeline [dict(typeLoadImage, backend_argsbackend_args),dict(typeGetBBoxCenterScale),dict(typeTopdownAffine, input_sizecodec[input_size]),dict(typePackPoseInputs)
]train_pipeline_stage2 [dict(typeLoadImage, backend_argsbackend_args),dict(typeGetBBoxCenterScale),dict(typeRandomFlip, directionhorizontal),dict(typeRandomHalfBody),dict(typeRandomBBoxTransform,shift_factor0.,scale_factor[0.75, 1.25],rotate_factor60),dict(typeTopdownAffine, input_sizecodec[input_size]),dict(typemmdet.YOLOXHSVRandomAug),dict(typeAlbumentation,transforms[dict(typeBlur, p0.1),dict(typeMedianBlur, p0.1),dict(typeCoarseDropout,max_holes1,max_height0.4,max_width0.4,min_holes1,min_height0.2,min_width0.2,p0.5),]),dict(typeGenerateTarget, encodercodec),dict(typePackPoseInputs)
]# data loaders
train_dataloader dict(batch_sizetrain_batch_size,num_workers2,persistent_workersTrue,samplerdict(typeDefaultSampler, shuffleTrue),datasetdict(typedataset_type,data_rootdata_root,metainfodataset_info,data_modedata_mode,ann_filetrainval.json,data_prefixdict(imgimages/),pipelinetrain_pipeline,))
val_dataloader dict(batch_sizeval_batch_size,num_workers2,persistent_workersTrue,drop_lastFalse,samplerdict(typeDefaultSampler, shuffleFalse, round_upFalse),datasetdict(typedataset_type,data_rootdata_root,metainfodataset_info,data_modedata_mode,ann_filetest.json,data_prefixdict(imgimages/),pipelineval_pipeline,))
test_dataloader val_dataloaderdefault_hooks {checkpoint: {save_best: PCK,rule: greater,max_keep_ckpts: 2},logger: {interval: 50}
}custom_hooks [dict(typeEMAHook,ema_typeExpMomentumEMA,momentum0.0002,update_buffersTrue,priority49),dict(typemmdet.PipelineSwitchHook,switch_epochmax_epochs - stage2_num_epochs,switch_pipelinetrain_pipeline_stage2)
]# evaluators
val_evaluator [dict(typeCocoMetric, ann_filedata_root test.json),dict(typePCKAccuracy),dict(typeAUC),dict(typeNME, norm_modekeypoint_distance, keypoint_indices[0, 1])
]test_evaluator val_evaluatorconfig ./configs/animal_2d_keypoint/cat_keypoint.py
with io.open(config, w, encodingutf-8) as f:f.write(custom_config)模型训练
使用训练脚本启动训练
!python tools/train.py {config}因为训练输出太长这里截取一段模型在测试集上最佳精度
08/06 19:15:56 - mmengine - INFO - Evaluating CocoMetric...
Loading and preparing results...
DONE (t0.07s)
creating index...
index created!
Running per image evaluation...
Evaluate annotation type *keypoints*
DONE (t0.57s).
Accumulating evaluation results...
DONE (t0.03s).Average Precision (AP) [ IoU0.50:0.95 | area all | maxDets 20 ] 0.943Average Precision (AP) [ IoU0.50 | area all | maxDets 20 ] 0.979Average Precision (AP) [ IoU0.75 | area all | maxDets 20 ] 0.969Average Precision (AP) [ IoU0.50:0.95 | areamedium | maxDets 20 ] -1.000Average Precision (AP) [ IoU0.50:0.95 | area large | maxDets 20 ] 0.944Average Recall (AR) [ IoU0.50:0.95 | area all | maxDets 20 ] 0.953Average Recall (AR) [ IoU0.50 | area all | maxDets 20 ] 0.987Average Recall (AR) [ IoU0.75 | area all | maxDets 20 ] 0.977Average Recall (AR) [ IoU0.50:0.95 | areamedium | maxDets 20 ] -1.000Average Recall (AR) [ IoU0.50:0.95 | area large | maxDets 20 ] 0.953
08/06 19:15:56 - mmengine - INFO - Evaluating PCKAccuracy (normalized by bbox_size)...
08/06 19:15:56 - mmengine - INFO - Evaluating AUC...
08/06 19:15:56 - mmengine - INFO - Evaluating NME...
08/06 19:15:57 - mmengine - INFO - Epoch(val) [60][47/47] coco/AP: 0.943453 coco/AP .5: 0.979424 coco/AP .75: 0.969202 coco/AP (M): -1.000000 coco/AP (L): 0.944082 coco/AR: 0.953471 coco/AR .5: 0.987316 coco/AR .75: 0.977303 coco/AR (M): -1.000000 coco/AR (L): 0.953471 PCK: 0.978045 AUC: 0.801710 NME: 0.121770 data_time: 0.101005 time: 0.435133
08/06 19:15:57 - mmengine - INFO - The previous best checkpoint /kaggle/working/mmpose/work_dir/best_PCK_epoch_55.pth is removed
08/06 19:16:01 - mmengine - INFO - The best checkpoint with 0.9780 PCK at 60 epoch is saved to best_PCK_epoch_60.pth.可以看到模型PCK达到了0.978AUC达到了0.8017mAP也都挺高说明模型效果非常不错
模型精简
mmpose提供模型精简脚本模型训练权重文件大小减少一半但不影响精度和推理将在验证集上表现最好的模型权重进行精简
import glob
ckpt_path glob.glob(./work_dir/best_PCK_*.pth)[0]
ckpt_sim ./work_dir/cat_pose_sim.pth
# 模型精简
!python tools/misc/publish_model.py \{ckpt_path} \{ckpt_sim}模型推理
这里和上面的模型推理使用相同的思路使用RTMDet模型进行目标检测使用我们自己训练的RTMPose模型进行姿态估计。不过pose2d参数是我们上面保存的配置文件./configs/animal_2d_keypoint/cat_keypoint.pypose2d_weights为最佳精度模型精简后的权重文件glob.glob(./work_dir/cat_pose_sim*.pth)[0]。
img_path /kaggle/input/cat-dataset/CAT_00/00000001_012.jpginferencer MMPoseInferencer(det_model /kaggle/working/mmdetection/configs/rtmdet/rtmdet_l_8xb32-300e_coco.py,det_weights checkpoint/rtmdet_l_8xb32-300e_coco_20220719_112030-5a0be7c4.pth,pose2d ./configs/animal_2d_keypoint/cat_keypoint.py,pose2d_weights glob.glob(./work_dir/cat_pose_sim*.pth)[0])result_generator inferencer(img_path, out_dir outputs, showFalse)
result next(result_generator)
display.clear_output()可视化训练结果
img_og mmcv.imread(img_path)
img_fuse mmcv.imread(outputs/visualizations/00000001_012.jpg)fig, axes plt.subplots(1, 2, figsize(15, 10))
axes[0].imshow(mmcv.bgr2rgb(img_og))
axes[0].set_title(Original Image)
axes[0].axis(off)axes[1].imshow(mmcv.bgr2rgb(img_fuse))
axes[1].set_title(Keypoint Image)
axes[1].axis(off)
plt.show()
