做网站推广有用吗,开发网站的申请怎么写,椒江网站建设公司,单页面的网站概述
目标检测(Object Detection)就是一种基于目标几何和统计特征的图像分割,它将目标的分割和识别合二为一,通俗点说就是给定一张图片要精确的定位到物体所在位置,并完成对物体类别的识别。其准确性和实时性是整个系统的一项重要能力。
R-CNN的全称是Region-CNN(区域卷积神经…概述
目标检测(Object Detection)就是一种基于目标几何和统计特征的图像分割,它将目标的分割和识别合二为一,通俗点说就是给定一张图片要精确的定位到物体所在位置,并完成对物体类别的识别。其准确性和实时性是整个系统的一项重要能力。
R-CNN的全称是Region-CNN(区域卷积神经网络),是第一个成功将深度学习应用到目标检测上的算法。R-CNN基于卷积神经网络(CNN),线性回归,和支持向量机(SVM)等算法,实现目标检测技术。
以下的代码和项目工程是引用他人的,此文只对其做一个简单的流程梳理。
这里先贴出工具脚本util.py的代码如下
# -*- coding: utf-8 -*-
date: 2020/2/29 下午7:31
file: util.py
author: zj
description:
import os
import numpy as np
import xmltodict
import torch
import matplotlib.pyplot as pltdef check_dir(data_dir):if not os.path.exists(data_dir):os.mkdir(data_dir)def parse_car_csv(csv_dir):csv_path os.path.join(csv_dir, car.csv)samples np.loadtxt(csv_path, dtypestr)return samplesdef parse_xml(xml_path):解析xml文件返回标注边界框坐标# print(xml_path)with open(xml_path, rb) as f:xml_dict xmltodict.parse(f)# print(xml_dict)bndboxs list()objects xml_dict[annotation][object]if isinstance(objects, list):for obj in objects:obj_name obj[name]difficult int(obj[difficult])if car.__eq__(obj_name) and difficult ! 1:bndbox obj[bndbox]bndboxs.append((int(bndbox[xmin]), int(bndbox[ymin]), int(bndbox[xmax]), int(bndbox[ymax])))elif isinstance(objects, dict):obj_name objects[name]difficult int(objects[difficult])if car.__eq__(obj_name) and difficult ! 1:bndbox objects[bndbox]bndboxs.append((int(bndbox[xmin]), int(bndbox[ymin]), int(bndbox[xmax]), int(bndbox[ymax])))else:passreturn np.array(bndboxs)def iou(pred_box, target_box):计算候选建议和标注边界框的IoU:param pred_box: 大小为[4]:param target_box: 大小为[N, 4]:return: [N]if len(target_box.shape) 1:target_box target_box[np.newaxis, :]xA np.maximum(pred_box[0], target_box[:, 0])yA np.maximum(pred_box[1], target_box[:, 1])xB np.minimum(pred_box[2], target_box[:, 2])yB np.minimum(pred_box[3], target_box[:, 3])# 计算交集面积intersection np.maximum(0.0, xB - xA) * np.maximum(0.0, yB - yA)# 计算两个边界框面积boxAArea (pred_box[2] - pred_box[0]) * (pred_box[3] - pred_box[1])boxBArea (target_box[:, 2] - target_box[:, 0]) * (target_box[:, 3] - target_box[:, 1])scores intersection / (boxAArea boxBArea - intersection)return scoresdef compute_ious(rects, bndboxs):iou_list list()for rect in rects:scores iou(rect, bndboxs)iou_list.append(max(scores))return iou_listdef save_model(model, model_save_path):# 保存最好的模型参数check_dir(./models)torch.save(model.state_dict(), model_save_path)def plot_loss(loss_list):x list(range(len(loss_list)))fg plt.figure()plt.plot(x, loss_list)plt.title(loss)plt.savefig(./loss.png)数据集准备
数据集下载
运行项目中的pascal_voc.py脚本,这个脚本是下载数据集。
# -*- coding: utf-8 -*-
date: 2020/2/29 下午2:51
file: pascal_voc.py
author: zj
description: 加载PASCAL VOC 2007数据集
import cv2
import numpy as np
from torchvision.datasets import VOCDetectionif __name__ __main__:下载PASCAL VOC数据集dataset VOCDetection(../../data, year2007, image_settrainval, downloadTrue)# img, target dataset.__getitem__(1000)# img np.array(img)# # print(target)# print(img.shape)# # cv2.imshow(img, img)# cv2.waitKey(0)从数据集中提取出car相关的数据
由于本文只针对汽车car进行目标检测,所以只需要car相关的数据。
执行pascal_voc_car.py脚本,脚本依次做了以下事:
①读取../../data/VOCdevkit/VOC2007/ImageSets/Main/car_train.txt文件和../../data/VOCdevkit/VOC2007/ImageSets/Main/car_val.txt文件
car_train.txt和car_val.txt文件的内容格式如下 ②然后将以上文件内容分别保存到../../data/voc_car/train/car.csv和../../data/voc_car/val/car.csv中
car.csv的内容格式如下 ③最后根据筛选出来的car的相关数据,从../../data/VOCdevkit/VOC2007/Annotations/中复制相关.xml文件到../../data/voc_car/train/Annotations/和../../data/voc_car/val/Annotations/,以及从../../data/VOCdevkit/VOC2007/JPEGImages/中复制相关.jpg文件到../../data/voc_car/train/JPEGImages/和../../data/voc_car/val/JPEGImages/
以下是pascal_voc_car.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/2/29 下午2:43
file: pascal_voc_car.py
author: zj
description: 从PASCAL VOC 2007数据集中抽取类别Car。保留1/10的数目
import os
import shutil
import random
import numpy as np
import xmltodict
from utils.util import check_dirsuffix_xml .xml
suffix_jpeg .jpgcar_train_path ../../data/VOCdevkit/VOC2007/ImageSets/Main/car_train.txt
car_val_path ../../data/VOCdevkit/VOC2007/ImageSets/Main/car_val.txtvoc_annotation_dir ../../data/VOCdevkit/VOC2007/Annotations/
voc_jpeg_dir ../../data/VOCdevkit/VOC2007/JPEGImages/car_root_dir ../../data/voc_car/def parse_train_val(data_path):提取指定类别图像samples []with open(data_path, r) as file:lines file.readlines()for line in lines:res line.strip().split( )if len(res) 3 and int(res[2]) 1:samples.append(res[0])return np.array(samples)def sample_train_val(samples):随机采样样本减少数据集个数留下1/10for name in [train, val]:dataset samples[name]length len(dataset)random_samples random.sample(range(length), int(length / 10))# print(random_samples)new_dataset dataset[random_samples]samples[name] new_datasetreturn samplesdef save_car(car_samples, data_root_dir, data_annotation_dir, data_jpeg_dir):保存类别Car的样本图片和标注文件for sample_name in car_samples:src_annotation_path os.path.join(voc_annotation_dir, sample_name suffix_xml)dst_annotation_path os.path.join(data_annotation_dir, sample_name suffix_xml)shutil.copyfile(src_annotation_path, dst_annotation_path)src_jpeg_path os.path.join(voc_jpeg_dir, sample_name suffix_jpeg)dst_jpeg_path os.path.join(data_jpeg_dir, sample_name suffix_jpeg)shutil.copyfile(src_jpeg_path, dst_jpeg_path)csv_path os.path.join(data_root_dir, car.csv)np.savetxt(csv_path, np.array(car_samples), fmt%s)if __name__ __main__:samples {train: parse_train_val(car_train_path), val: parse_train_val(car_val_path)}print(samples)# samples sample_train_val(samples)# print(samples)check_dir(car_root_dir)for name in [train, val]:data_root_dir os.path.join(car_root_dir, name)data_annotation_dir os.path.join(data_root_dir, Annotations)data_jpeg_dir os.path.join(data_root_dir, JPEGImages)check_dir(data_root_dir)check_dir(data_annotation_dir)check_dir(data_jpeg_dir)save_car(samples[name], data_root_dir, data_annotation_dir, data_jpeg_dir)print(done)卷积神经网络微调模型
准备微调数据正负样本集
执行create_finetune_data.py脚本,这个脚本主要做了以下事
①把../../data/voc_car/train/JPEGImages/和../../data/voc_car/val/JPEGImages/中的.jpg文件复制到../../data/finetune_car/train/JPEGImages/和../../data/finetune_car/val/JPEGImages/,然后又把../../data/voc_car/train/car.csv和../../data/voc_car/val/car.csv分别复制到../../data/finetune_car/train/car.csv和../../data/finetune_car/val/car.csv
②根据../../data/finetune_car/train/car.csv和../../data/finetune_car/val/car.csv文件内容分别读取../../data/finetune_car/train/JPEGImages/和../../data/finetune_car/val/JPEGImages/中的图片,并传入parse_annotation_jpeg方法
③parse_annotation_jpeg方法中,先获取候选框rects,然后从.xml文件中获取标注框bndboxs,接着计算候选框和标注框的IoU得到iou_list,遍历iou_list,选出IoU≥0.5的作为正样本,0IoU0.5的作为负样本,且为了进一步限制负样本的数量,其大小必须大于最大标注框的面积的1/5。最后得到微调数据集正样本集positive_list和负样本集negative_list,并将其保存到../../data/finetune_car/train/Annotations/和../../data/finetune_car/val/Annotations/文件夹中,其文件名格式如下(_0表示负样本,_1表示正样本) 其文件内容格式如下 以下是create_finetune_data.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/2/29 下午7:22
file: create_finetune_data.py
author: zj
description: 创建微调数据集
import time
import shutil
import numpy as np
import cv2
import os
import selectivesearch
from utils.util import check_dir
from utils.util import parse_car_csv
from utils.util import parse_xml
from utils.util import compute_ious# train
# positive num: 66517
# negatie num: 464340
# val
# positive num: 64712
# negative num: 415134def parse_annotation_jpeg(annotation_path, jpeg_path, gs):获取正负样本注忽略属性difficult为True的标注边界框正样本候选建议与标注边界框IoU大于等于0.5负样本IoU大于0,小于0.5。为了进一步限制负样本数目其大小必须大于标注框的1/5img cv2.imread(jpeg_path)selectivesearch.config(gs, img, strategyq)# 计算候选建议rects selectivesearch.get_rects(gs)# 获取标注边界框bndboxs parse_xml(annotation_path)# 标注框大小maximum_bndbox_size 0for bndbox in bndboxs:xmin, ymin, xmax, ymax bndboxbndbox_size (ymax - ymin) * (xmax - xmin)if bndbox_size maximum_bndbox_size:maximum_bndbox_size bndbox_size# 获取候选建议和标注边界框的IoUiou_list compute_ious(rects, bndboxs)positive_list list()negative_list list()for i in range(len(iou_list)):xmin, ymin, xmax, ymax rects[i]rect_size (ymax - ymin) * (xmax - xmin)iou_score iou_list[i]if iou_list[i] 0.5:# 正样本positive_list.append(rects[i])if 0 iou_list[i] 0.5 and rect_size maximum_bndbox_size / 5.0:# 负样本negative_list.append(rects[i])else:passreturn positive_list, negative_listif __name__ __main__:car_root_dir ../../data/voc_car/finetune_root_dir ../../data/finetune_car/check_dir(finetune_root_dir)gs selectivesearch.get_selective_search()for name in [train, val]:src_root_dir os.path.join(car_root_dir, name)src_annotation_dir os.path.join(src_root_dir, Annotations)src_jpeg_dir os.path.join(src_root_dir, JPEGImages)dst_root_dir os.path.join(finetune_root_dir, name)dst_annotation_dir os.path.join(dst_root_dir, Annotations)dst_jpeg_dir os.path.join(dst_root_dir, JPEGImages)check_dir(dst_root_dir)check_dir(dst_annotation_dir)check_dir(dst_jpeg_dir)total_num_positive 0total_num_negative 0samples parse_car_csv(src_root_dir)# 复制csv文件src_csv_path os.path.join(src_root_dir, car.csv)dst_csv_path os.path.join(dst_root_dir, car.csv)shutil.copyfile(src_csv_path, dst_csv_path)for sample_name in samples:since time.time()src_annotation_path os.path.join(src_annotation_dir, sample_name .xml)src_jpeg_path os.path.join(src_jpeg_dir, sample_name .jpg)# 获取正负样本positive_list, negative_list parse_annotation_jpeg(src_annotation_path, src_jpeg_path, gs)total_num_positive len(positive_list)total_num_negative len(negative_list)dst_annotation_positive_path os.path.join(dst_annotation_dir, sample_name _1 .csv)dst_annotation_negative_path os.path.join(dst_annotation_dir, sample_name _0 .csv)dst_jpeg_path os.path.join(dst_jpeg_dir, sample_name .jpg)# 保存图片shutil.copyfile(src_jpeg_path, dst_jpeg_path)# 保存正负样本标注np.savetxt(dst_annotation_positive_path, np.array(positive_list), fmt%d, delimiter )np.savetxt(dst_annotation_negative_path, np.array(negative_list), fmt%d, delimiter )time_elapsed time.time() - sinceprint(parse {}.png in {:.0f}m {:.0f}s.format(sample_name, time_elapsed // 60, time_elapsed % 60))print(%s positive num: %d % (name, total_num_positive))print(%s negative num: %d % (name, total_num_negative))print(done)自定义微调数据集类
custom_finetune_dataset.py,该脚本不用主动执行,在训练微调模型的时候,自然会调用到,以下只说这个脚本做了什么事
①CustomFinetuneDataset类继承自Dataset
②__init__时读取../../data/finetune_car/train/JPEGImages/或../../data/finetune_car/val/JPEGImages/文件夹中的图片,读取../../data/finetune_car/train/Annotations/或../../data/finetune_car/val/Annotations/中的正负样本集,记录正样本总数self.total_positive_num,负样本总数self.total_negative_num,正样本候选框positive_rects,负样本候选框negative_rects
③__getitem__方法传入index,如果index小于正样本总数self.total_positive_num,那么返回对应正样本的图片和标签(1),否则返回对应负样本的图片和标签(0)。
以下是custom_finetune_dataset.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/3/3 下午7:06
file: custom_finetune_dataset.py
author: zj
description: 自定义微调数据类
import numpy as np
import os
import cv2
from PIL import Image
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
import torchvision.transforms as transformsfrom utils.util import parse_car_csvclass CustomFinetuneDataset(Dataset):def __init__(self, root_dir, transformNone):samples parse_car_csv(root_dir)jpeg_images [cv2.imread(os.path.join(root_dir, JPEGImages, sample_name .jpg))for sample_name in samples]positive_annotations [os.path.join(root_dir, Annotations, sample_name _1.csv)for sample_name in samples]negative_annotations [os.path.join(root_dir, Annotations, sample_name _0.csv)for sample_name in samples]# 边界框大小positive_sizes list()negative_sizes list()# 边界框坐标positive_rects list()negative_rects list()for annotation_path in positive_annotations:rects np.loadtxt(annotation_path, dtypeint, delimiter )# 存在文件为空或者文件中仅有单行数据if len(rects.shape) 1:# 是否为单行if rects.shape[0] 4:positive_rects.append(rects)positive_sizes.append(1)else:positive_sizes.append(0)else:positive_rects.extend(rects)positive_sizes.append(len(rects))for annotation_path in negative_annotations:rects np.loadtxt(annotation_path, dtypeint, delimiter )# 和正样本规则一样if len(rects.shape) 1:if rects.shape[0] 4:negative_rects.append(rects)negative_sizes.append(1)else:positive_sizes.append(0)else:negative_rects.extend(rects)negative_sizes.append(len(rects))print(positive_rects)self.transform transformself.jpeg_images jpeg_imagesself.positive_sizes positive_sizesself.negative_sizes negative_sizesself.positive_rects positive_rectsself.negative_rects negative_rectsself.total_positive_num int(np.sum(positive_sizes))self.total_negative_num int(np.sum(negative_sizes))def __getitem__(self, index: int):# 定位下标所属图像image_id len(self.jpeg_images) - 1if index self.total_positive_num:# 正样本target 1xmin, ymin, xmax, ymax self.positive_rects[index]# 寻找所属图像for i in range(len(self.positive_sizes) - 1):if np.sum(self.positive_sizes[:i]) index np.sum(self.positive_sizes[:(i 1)]):image_id ibreakimage self.jpeg_images[image_id][ymin:ymax, xmin:xmax]else:# 负样本target 0idx index - self.total_positive_numxmin, ymin, xmax, ymax self.negative_rects[idx]# 寻找所属图像for i in range(len(self.negative_sizes) - 1):if np.sum(self.negative_sizes[:i]) idx np.sum(self.negative_sizes[:(i 1)]):image_id ibreakimage self.jpeg_images[image_id][ymin:ymax, xmin:xmax]# print(index: %d image_id: %d target: %d image.shape: %s [xmin, ymin, xmax, ymax]: [%d, %d, %d, %d] %# (index, image_id, target, str(image.shape), xmin, ymin, xmax, ymax))if self.transform:image self.transform(image)return image, targetdef __len__(self) - int:return self.total_positive_num self.total_negative_numdef get_positive_num(self) - int:return self.total_positive_numdef get_negative_num(self) - int:return self.total_negative_numdef test(idx):root_dir ../../data/finetune_car/traintrain_data_set CustomFinetuneDataset(root_dir)print(positive num: %d % train_data_set.get_positive_num())print(negative num: %d % train_data_set.get_negative_num())print(total num: %d % train_data_set.__len__())# 测试id3/66516/66517/530856image, target train_data_set.__getitem__(idx)print(target: %d % target)image Image.fromarray(image)print(image)print(type(image))# cv2.imshow(image, image)# cv2.waitKey(0)def test2():root_dir ../../data/finetune_car/traintransform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])train_data_set CustomFinetuneDataset(root_dir, transformtransform)image, target train_data_set.__getitem__(530856)print(target: %d % target)print(image.shape: str(image.shape))def test3():root_dir ../../data/finetune_car/traintransform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])train_data_set CustomFinetuneDataset(root_dir, transformtransform)data_loader DataLoader(train_data_set, batch_size128, num_workers8, drop_lastTrue)inputs, targets next(data_loader.__iter__())print(targets)print(inputs.shape)if __name__ __main__:# test(159622)# test(4051)test3()自定义批量采样器类
custom_batch_sampler.py,该脚本也不用主动执行,在训练微调模型的时候,自然会调用到,以下只说这个脚本做了什么事
①CustomBatchSampler类继承自(Sampler)
②__init__时通过传入的正样本总数num_positive和负样本总数num_negative得出一个列表self.idx_list,并结合传入的单次正样本数batch_positive和单次负样本数batch_negative算出可迭代次数self.num_iter
③__iter__方法中循环self.num_iter次,每次循环中会对正样本随机采集self.batch_positive次index,以及对负样本随机采集self.batch_negative次index,然后打乱存入sampler_list,最后返回一个迭代器iter(sampler)
以下是custom_batch_sampler.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/3/3 下午7:38
file: custom_batch_sampler.py
author: zj
description: 自定义采样器
import numpy as np
import random
from torch.utils.data import Sampler
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
from utils.data.custom_finetune_dataset import CustomFinetuneDatasetclass CustomBatchSampler(Sampler):def __init__(self, num_positive, num_negative, batch_positive, batch_negative) - None:2分类数据集每次批量处理其中batch_positive个正样本batch_negative个负样本param num_positive: 正样本数目param num_negative: 负样本数目param batch_positive: 单次正样本数param batch_negative: 单次负样本数self.num_positive num_positiveself.num_negative num_negativeself.batch_positive batch_positiveself.batch_negative batch_negativelength num_positive num_negativeself.idx_list list(range(length))self.batch batch_negative batch_positiveself.num_iter length // self.batchdef __iter__(self):sampler_list list()for i in range(self.num_iter):tmp np.concatenate((random.sample(self.idx_list[:self.num_positive], self.batch_positive),random.sample(self.idx_list[self.num_positive:], self.batch_negative)))random.shuffle(tmp)sampler_list.extend(tmp)return iter(sampler_list)def __len__(self) - int:return self.num_iter * self.batchdef get_num_batch(self) - int:return self.num_iterdef test():root_dir ../../data/finetune_car/traintrain_data_set CustomFinetuneDataset(root_dir)train_sampler CustomBatchSampler(train_data_set.get_positive_num(), train_data_set.get_negative_num(), 32, 96)print(sampler len: %d % train_sampler.__len__())print(sampler batch num: %d % train_sampler.get_num_batch())first_idx_list list(train_sampler.__iter__())[:128]print(first_idx_list)# 单次批量中正样本个数print(positive batch: %d % np.sum(np.array(first_idx_list) 66517))def test2():root_dir ../../data/finetune_car/traintransform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])train_data_set CustomFinetuneDataset(root_dir, transformtransform)train_sampler CustomBatchSampler(train_data_set.get_positive_num(), train_data_set.get_negative_num(), 32, 96)data_loader DataLoader(train_data_set, batch_size128, samplertrain_sampler, num_workers8, drop_lastTrue)inputs, targets next(data_loader.__iter__())print(targets)print(inputs.shape)if __name__ __main__:test()训练微调模型
执行finetune.py脚本
①调用custom_finetune_dataset.py脚本和custom_batch_sampler.py脚本,得到训练数据data_loaders
②使用预训练模型AlexNet网络模型,修改分类器对象classifier的输出为2类(1类是car,一类是背景),然后定义损失函数为交叉熵损失函数,优化函数为SGD,学习率更新器为StepLR,然后开始训练,保存准确率最高的权重到models/alexnet_car.pth
以下是finetune.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/3/1 上午9:54
file: finetune.py
author: zj
description:
import os
import copy
import time
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
import torchvision.models as models
from torchvision.models import AlexNet_Weightsfrom utils.data.custom_finetune_dataset import CustomFinetuneDataset
from utils.data.custom_batch_sampler import CustomBatchSampler
from utils.util import check_dirdef load_data(data_root_dir):transform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.RandomHorizontalFlip(),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])data_loaders {}data_sizes {}for name in [train, val]:data_dir os.path.join(data_root_dir, name)data_set CustomFinetuneDataset(data_dir, transformtransform)data_sampler CustomBatchSampler(data_set.get_positive_num(), data_set.get_negative_num(), 32, 96)data_loader DataLoader(data_set, batch_size128, samplerdata_sampler, num_workers8, drop_lastTrue)data_loaders[name] data_loaderdata_sizes[name] data_sampler.__len__()return data_loaders, data_sizesdef train_model(data_loaders, model, criterion, optimizer, lr_scheduler, num_epochs25, deviceNone):since time.time()best_model_weights copy.deepcopy(model.state_dict())best_acc 0.0for epoch in range(num_epochs):print(Epoch {}/{}.format(epoch, num_epochs - 1))print(- * 10)# Each epoch has a training and validation phasefor phase in [train, val]:if phase train:model.train() # Set model to training modeelse:model.eval() # Set model to evaluate moderunning_loss 0.0running_corrects 0# Iterate over data.for inputs, labels in data_loaders[phase]:inputs inputs.to(device)labels labels.to(device)# zero the parameter gradientsoptimizer.zero_grad()# forward# track history if only in trainwith torch.set_grad_enabled(phase train):outputs model(inputs)_, preds torch.max(outputs, 1)loss criterion(outputs, labels)# backward optimize only if in training phaseif phase train:loss.backward()optimizer.step()# statisticsrunning_loss loss.item() * inputs.size(0)running_corrects torch.sum(preds labels.data)if phase train:lr_scheduler.step()epoch_loss running_loss / data_sizes[phase]epoch_acc running_corrects.double() / data_sizes[phase]print({} Loss: {:.4f} Acc: {:.4f}.format(phase, epoch_loss, epoch_acc))# deep copy the modelif phase val and epoch_acc best_acc:best_acc epoch_accbest_model_weights copy.deepcopy(model.state_dict())print()time_elapsed time.time() - sinceprint(Training complete in {:.0f}m {:.0f}s.format(time_elapsed // 60, time_elapsed % 60))print(Best val Acc: {:4f}.format(best_acc))# load best model weightsmodel.load_state_dict(best_model_weights)return modelif __name__ __main__:device torch.device(cuda:0 if torch.cuda.is_available() else cpu)data_loaders, data_sizes load_data(./data/finetune_car)model models.alexnet(weightsAlexNet_Weights.IMAGENET1K_V1)# print(model)num_features model.classifier[6].in_featuresmodel.classifier[6] nn.Linear(num_features, 2)# print(model)model model.to(device)criterion nn.CrossEntropyLoss()optimizer optim.SGD(model.parameters(), lr1e-3, momentum0.9)lr_scheduler optim.lr_scheduler.StepLR(optimizer, step_size7, gamma0.1)best_model train_model(data_loaders, model, criterion, optimizer, lr_scheduler, devicedevice, num_epochs25)# 保存最好的模型参数check_dir(./models)torch.save(best_model.state_dict(), models/alexnet_car.pth)分类器训练
准备分类器数据集
执行create_classifier_data.py脚本
①把../../data/finetune_car/train/JPEGImages/和../../data/finetune_car/val/JPEGImages/中的.jpg文件复制到../../data/classifier_car/train/JPEGImages/和../../data/classifier_car/val/JPEGImages/,然后又把../../data/finetune_car/train/car.csv和../../data/finetune_car/val/car.csv分别复制到../../data/classifier_car/train/car.csv和../../data/classifier_car/val/car.csv
②根据../../data/classifier_car/train/car.csv和../../data/classifier_car/val/car.csv文件内容分别读取../../data/classifier_car/train/JPEGImages/和../../data/classifier_car/val/JPEGImages/中的图片,并传入parse_annotation_jpeg方法
③parse_annotation_jpeg方法中,先获取候选框rects,然后从.xml文件中获取标注框bndboxs,接着计算候选框和标注框的IoU得到iou_list,遍历iou_list,选出0IoU≤0.3的候选框作为负样本,且为了进一步限制负样本的数量,其大小必须大于最大标注框的面积的1/5。最后得到分类器数据集正样本集positive_list(此处标注框bndboxs作为正样本集)和负样本集negative_list,并将其保存到../../data/classifier_car/train/Annotations/和../../data/classifier_car/val/Annotations/文件夹中,其文件名格式如下(_0表示负样本,_1表示正样本)
以下是create_classifier_data.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/3/1 下午7:17
file: create_classifier_data.py
author: zj
description: 创建分类器数据集
import random
import numpy as np
import shutil
import time
import cv2
import os
import xmltodict
import selectivesearch
from utils.util import check_dir
from utils.util import parse_car_csv
from utils.util import parse_xml
from utils.util import iou
from utils.util import compute_ious# train
# positive num: 625
# negative num: 366028
# val
# positive num: 625
# negative num: 321474def parse_annotation_jpeg(annotation_path, jpeg_path, gs):获取正负样本注忽略属性difficult为True的标注边界框正样本标注边界框负样本IoU大于0小于等于0.3。为了进一步限制负样本数目其大小必须大于标注框的1/5img cv2.imread(jpeg_path)selectivesearch.config(gs, img, strategyq)# 计算候选建议rects selectivesearch.get_rects(gs)# 获取标注边界框bndboxs parse_xml(annotation_path)# 标注框大小maximum_bndbox_size 0for bndbox in bndboxs:xmin, ymin, xmax, ymax bndboxbndbox_size (ymax - ymin) * (xmax - xmin)if bndbox_size maximum_bndbox_size:maximum_bndbox_size bndbox_size# 获取候选建议和标注边界框的IoUiou_list compute_ious(rects, bndboxs)positive_list list()negative_list list()for i in range(len(iou_list)):xmin, ymin, xmax, ymax rects[i]rect_size (ymax - ymin) * (xmax - xmin)iou_score iou_list[i]if 0 iou_score 0.3 and rect_size maximum_bndbox_size / 5.0:# 负样本negative_list.append(rects[i])else:passreturn bndboxs, negative_listif __name__ __main__:car_root_dir ../../data/voc_car/classifier_root_dir ../../data/classifier_car/check_dir(classifier_root_dir)gs selectivesearch.get_selective_search()for name in [train, val]:src_root_dir os.path.join(car_root_dir, name)src_annotation_dir os.path.join(src_root_dir, Annotations)src_jpeg_dir os.path.join(src_root_dir, JPEGImages)dst_root_dir os.path.join(classifier_root_dir, name)dst_annotation_dir os.path.join(dst_root_dir, Annotations)dst_jpeg_dir os.path.join(dst_root_dir, JPEGImages)check_dir(dst_root_dir)check_dir(dst_annotation_dir)check_dir(dst_jpeg_dir)total_num_positive 0total_num_negative 0samples parse_car_csv(src_root_dir)# 复制csv文件src_csv_path os.path.join(src_root_dir, car.csv)dst_csv_path os.path.join(dst_root_dir, car.csv)shutil.copyfile(src_csv_path, dst_csv_path)for sample_name in samples:since time.time()src_annotation_path os.path.join(src_annotation_dir, sample_name .xml)src_jpeg_path os.path.join(src_jpeg_dir, sample_name .jpg)# 获取正负样本positive_list, negative_list parse_annotation_jpeg(src_annotation_path, src_jpeg_path, gs)total_num_positive len(positive_list)total_num_negative len(negative_list)dst_annotation_positive_path os.path.join(dst_annotation_dir, sample_name _1 .csv)dst_annotation_negative_path os.path.join(dst_annotation_dir, sample_name _0 .csv)dst_jpeg_path os.path.join(dst_jpeg_dir, sample_name .jpg)# 保存图片shutil.copyfile(src_jpeg_path, dst_jpeg_path)# 保存正负样本标注np.savetxt(dst_annotation_positive_path, np.array(positive_list), fmt%d, delimiter )np.savetxt(dst_annotation_negative_path, np.array(negative_list), fmt%d, delimiter )time_elapsed time.time() - sinceprint(parse {}.png in {:.0f}m {:.0f}s.format(sample_name, time_elapsed // 60, time_elapsed % 60))print(%s positive num: %d % (name, total_num_positive))print(%s negative num: %d % (name, total_num_negative))print(done)自定义分类器数据集类
custom_classifier_dataset.py,该脚本不用主动执行,在训练分类器模型的时候,自然会调用到,以下只说这个脚本做了什么事
①CustomClassifierDataset类继承自Dataset
②__init__时读取../../data/classifier_car/train/JPEGImages/或../../data/classifier_car/val/JPEGImages/文件夹中的图片,读取../../data/classifier_car/train/Annotations/或../../data/classifier_car/val/Annotations/中的正负样本集,记录正样本列表self.positive_list,负样本总数self.negative_list,正样本候选框positive_rects,负样本候选框negative_rects
③__getitem__方法传入index,如果index小于正样本总数len(self.positive_list),那么返回对应正样本的图片和标签(1),否则返回对应负样本的图片和标签(0)。
以下是custom_classifier_dataset.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/3/4 下午4:00
file: custom_classifier_dataset.py
author: zj
description: 分类器数据集类可进行正负样本集替换适用于hard negative mining操作
import numpy as np
import os
import cv2
from PIL import Image
from torch.utils.data import Dataset
from torch.utils.data import DataLoader
import torchvision.transforms as transformsfrom utils.util import parse_car_csvclass CustomClassifierDataset(Dataset):def __init__(self, root_dir, transformNone):samples parse_car_csv(root_dir)jpeg_images list()positive_list list()negative_list list()for idx in range(len(samples)):sample_name samples[idx]jpeg_images.append(cv2.imread(os.path.join(root_dir, JPEGImages, sample_name .jpg)))positive_annotation_path os.path.join(root_dir, Annotations, sample_name _1.csv)positive_annotations np.loadtxt(positive_annotation_path, dtypeint, delimiter )# 考虑csv文件为空或者仅包含单个标注框if len(positive_annotations.shape) 1:# 单个标注框坐标if positive_annotations.shape[0] 4:positive_dict dict()positive_dict[rect] positive_annotationspositive_dict[image_id] idx# positive_dict[image_name] sample_namepositive_list.append(positive_dict)else:for positive_annotation in positive_annotations:positive_dict dict()positive_dict[rect] positive_annotationpositive_dict[image_id] idx# positive_dict[image_name] sample_namepositive_list.append(positive_dict)negative_annotation_path os.path.join(root_dir, Annotations, sample_name _0.csv)negative_annotations np.loadtxt(negative_annotation_path, dtypeint, delimiter )# 考虑csv文件为空或者仅包含单个标注框if len(negative_annotations.shape) 1:# 单个标注框坐标if negative_annotations.shape[0] 4:negative_dict dict()negative_dict[rect] negative_annotationsnegative_dict[image_id] idx# negative_dict[image_name] sample_namenegative_list.append(negative_dict)else:for negative_annotation in negative_annotations:negative_dict dict()negative_dict[rect] negative_annotationnegative_dict[image_id] idx# negative_dict[image_name] sample_namenegative_list.append(negative_dict)self.transform transformself.jpeg_images jpeg_imagesself.positive_list positive_listself.negative_list negative_listdef __getitem__(self, index: int):# 定位下标所属图像if index len(self.positive_list):# 正样本target 1positive_dict self.positive_list[index]xmin, ymin, xmax, ymax positive_dict[rect]image_id positive_dict[image_id]image self.jpeg_images[image_id][ymin:ymax, xmin:xmax]cache_dict positive_dictelse:# 负样本target 0idx index - len(self.positive_list)negative_dict self.negative_list[idx]xmin, ymin, xmax, ymax negative_dict[rect]image_id negative_dict[image_id]image self.jpeg_images[image_id][ymin:ymax, xmin:xmax]cache_dict negative_dict# print(index: %d image_id: %d target: %d image.shape: %s [xmin, ymin, xmax, ymax]: [%d, %d, %d, %d] %# (index, image_id, target, str(image.shape), xmin, ymin, xmax, ymax))if self.transform:image self.transform(image)return image, target, cache_dictdef __len__(self) - int:return len(self.positive_list) len(self.negative_list)def get_transform(self):return self.transformdef get_jpeg_images(self) - list:return self.jpeg_imagesdef get_positive_num(self) - int:return len(self.positive_list)def get_negative_num(self) - int:return len(self.negative_list)def get_positives(self) - list:return self.positive_listdef get_negatives(self) - list:return self.negative_list# 用于hard negative mining# 替换负样本def set_negative_list(self, negative_list):self.negative_list negative_listdef test(idx):root_dir ../../data/classifier_car/valtrain_data_set CustomClassifierDataset(root_dir)print(positive num: %d % train_data_set.get_positive_num())print(negative num: %d % train_data_set.get_negative_num())print(total num: %d % train_data_set.__len__())# 测试id3/66516/66517/530856image, target, cache_dict train_data_set.__getitem__(idx)print(target: %d % target)print(dict: str(cache_dict))image Image.fromarray(image)print(image)print(type(image))# cv2.imshow(image, image)# cv2.waitKey(0)def test2():root_dir ../../data/classifier_car/traintransform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])train_data_set CustomClassifierDataset(root_dir, transformtransform)image, target, cache_dict train_data_set.__getitem__(230856)print(target: %d % target)print(dict: str(cache_dict))print(image.shape: str(image.shape))def test3():root_dir ../../data/classifier_car/traintransform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])train_data_set CustomClassifierDataset(root_dir, transformtransform)data_loader DataLoader(train_data_set, batch_size128, num_workers8, drop_lastTrue)inputs, targets, cache_dicts next(data_loader.__iter__())print(targets)print(inputs.shape)if __name__ __main__:# test(159622)# test(4051)test(24768)# test2()# test3()自定义批量采样器类
同卷积神经网络微调模型中的自定义批量采样器类,在训练分类器模型的时候,自然会调用到
训练分类器
执行linear_svm.py脚本
①调用custom_classifier_dataset.py脚本和custom_batch_sampler.py脚本,得到训练数据data_loaders
②使用AlexNet网络模型,修改分类器对象classifier的输出为2类(1类是car,一类是背景),加载之前微调训练的权重alexnet_car.pth,并设置参数冻结,然后再添加一个全连接层作为svm分类器,定义损失函数为折页损失函数,优化函数为SGD,学习率更新器为StepLR,然后开始训练,保存准确率最高的权重到models/best_linear_svm_alexnet_car.pth
以下是linear_svm.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/3/1 下午2:38
file: linear_svm.py
author: zj
description:
import time
import copy
import os
import random
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
from torchvision.models import alexnetfrom utils.data.custom_classifier_dataset import CustomClassifierDataset
from utils.data.custom_hard_negative_mining_dataset import CustomHardNegativeMiningDataset
from utils.data.custom_batch_sampler import CustomBatchSampler
from utils.util import check_dir
from utils.util import save_modelbatch_positive 32
batch_negative 96
batch_total 128def load_data(data_root_dir):transform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.RandomHorizontalFlip(),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])data_loaders {}data_sizes {}remain_negative_list list()for name in [train, val]:data_dir os.path.join(data_root_dir, name)data_set CustomClassifierDataset(data_dir, transformtransform)if name train:使用hard negative mining方式初始正负样本比例为1:1。由于正样本数远小于负样本所以以正样本数为基准在负样本集中随机提取同样数目负样本作为初始负样本集positive_list data_set.get_positives()negative_list data_set.get_negatives()init_negative_idxs random.sample(range(len(negative_list)), len(positive_list))init_negative_list [negative_list[idx] for idx in range(len(negative_list)) if idx in init_negative_idxs]remain_negative_list [negative_list[idx] for idx in range(len(negative_list))if idx not in init_negative_idxs]data_set.set_negative_list(init_negative_list)data_loaders[remain] remain_negative_listsampler CustomBatchSampler(data_set.get_positive_num(), data_set.get_negative_num(),batch_positive, batch_negative)data_loader DataLoader(data_set, batch_sizebatch_total, samplersampler, num_workers8, drop_lastTrue)data_loaders[name] data_loaderdata_sizes[name] len(sampler)return data_loaders, data_sizesdef hinge_loss(outputs, labels):折页损失计算:param outputs: 大小为(N, num_classes):param labels: 大小为(N):return: 损失值num_labels len(labels)corrects outputs[range(num_labels), labels].unsqueeze(0).T# 最大间隔margin 1.0margins outputs - corrects marginloss torch.sum(torch.max(margins, 1)[0]) / len(labels)# # 正则化强度# reg 1e-3# loss reg * torch.sum(weight ** 2)return lossdef add_hard_negatives(hard_negative_list, negative_list, add_negative_list):for item in hard_negative_list:if len(add_negative_list) 0:# 第一次添加负样本negative_list.append(item)add_negative_list.append(list(item[rect]))if list(item[rect]) not in add_negative_list:negative_list.append(item)add_negative_list.append(list(item[rect]))def get_hard_negatives(preds, cache_dicts):fp_mask preds 1tn_mask preds 0fp_rects cache_dicts[rect][fp_mask].numpy()fp_image_ids cache_dicts[image_id][fp_mask].numpy()tn_rects cache_dicts[rect][tn_mask].numpy()tn_image_ids cache_dicts[image_id][tn_mask].numpy()hard_negative_list [{rect: fp_rects[idx], image_id: fp_image_ids[idx]} for idx in range(len(fp_rects))]easy_negatie_list [{rect: tn_rects[idx], image_id: tn_image_ids[idx]} for idx in range(len(tn_rects))]return hard_negative_list, easy_negatie_listdef train_model(data_loaders, model, criterion, optimizer, lr_scheduler, num_epochs25, deviceNone):since time.time()best_model_weights copy.deepcopy(model.state_dict())best_acc 0.0for epoch in range(num_epochs):print(Epoch {}/{}.format(epoch, num_epochs - 1))print(- * 10)# Each epoch has a training and validation phasefor phase in [train, val]:if phase train:model.train() # Set model to training modeelse:model.eval() # Set model to evaluate moderunning_loss 0.0running_corrects 0# 输出正负样本数data_set data_loaders[phase].datasetprint({} - positive_num: {} - negative_num: {} - data size: {}.format(phase, data_set.get_positive_num(), data_set.get_negative_num(), data_sizes[phase]))# Iterate over data.for inputs, labels, cache_dicts in data_loaders[phase]:inputs inputs.to(device)labels labels.to(device)# zero the parameter gradientsoptimizer.zero_grad()# forward# track history if only in trainwith torch.set_grad_enabled(phase train):outputs model(inputs)# print(outputs.shape)_, preds torch.max(outputs, 1)loss criterion(outputs, labels)# backward optimize only if in training phaseif phase train:loss.backward()optimizer.step()# statisticsrunning_loss loss.item() * inputs.size(0)running_corrects torch.sum(preds labels.data)if phase train:lr_scheduler.step()epoch_loss running_loss / data_sizes[phase]epoch_acc running_corrects.double() / data_sizes[phase]print({} Loss: {:.4f} Acc: {:.4f}.format(phase, epoch_loss, epoch_acc))# deep copy the modelif phase val and epoch_acc best_acc:best_acc epoch_accbest_model_weights copy.deepcopy(model.state_dict())# 每一轮训练完成后测试剩余负样本集进行hard negative miningtrain_dataset data_loaders[train].datasetremain_negative_list data_loaders[remain]jpeg_images train_dataset.get_jpeg_images()transform train_dataset.get_transform()with torch.set_grad_enabled(False):remain_dataset CustomHardNegativeMiningDataset(remain_negative_list, jpeg_images, transformtransform)remain_data_loader DataLoader(remain_dataset, batch_sizebatch_total, num_workers8, drop_lastTrue)# 获取训练数据集的负样本集negative_list train_dataset.get_negatives()# 记录后续增加的负样本add_negative_list data_loaders.get(add_negative, [])running_corrects 0# Iterate over data.for inputs, labels, cache_dicts in remain_data_loader:inputs inputs.to(device)labels labels.to(device)# zero the parameter gradientsoptimizer.zero_grad()outputs model(inputs)# print(outputs.shape)_, preds torch.max(outputs, 1)running_corrects torch.sum(preds labels.data)hard_negative_list, easy_neagtive_list get_hard_negatives(preds.cpu().numpy(), cache_dicts)add_hard_negatives(hard_negative_list, negative_list, add_negative_list)remain_acc running_corrects.double() / len(remain_negative_list)print(remain negative size: {}, acc: {:.4f}.format(len(remain_negative_list), remain_acc))# 训练完成后重置负样本进行hard negatives miningtrain_dataset.set_negative_list(negative_list)tmp_sampler CustomBatchSampler(train_dataset.get_positive_num(), train_dataset.get_negative_num(),batch_positive, batch_negative)data_loaders[train] DataLoader(train_dataset, batch_sizebatch_total, samplertmp_sampler,num_workers8, drop_lastTrue)data_loaders[add_negative] add_negative_list# 重置数据集大小data_sizes[train] len(tmp_sampler)# 每训练一轮就保存save_model(model, models/linear_svm_alexnet_car_%d.pth % epoch)time_elapsed time.time() - sinceprint(Training complete in {:.0f}m {:.0f}s.format(time_elapsed // 60, time_elapsed % 60))print(Best val Acc: {:4f}.format(best_acc))# load best model weightsmodel.load_state_dict(best_model_weights)return modelif __name__ __main__:device torch.device(cuda:0 if torch.cuda.is_available() else cpu)# device cpudata_loaders, data_sizes load_data(./data/classifier_car)# 加载CNN模型model_path ./models/alexnet_car.pthmodel alexnet()num_classes 2num_features model.classifier[6].in_featuresmodel.classifier[6] nn.Linear(num_features, num_classes)model.load_state_dict(torch.load(model_path))model.eval()# 固定特征提取for param in model.parameters():param.requires_grad False# 创建SVM分类器model.classifier[6] nn.Linear(num_features, num_classes)# print(model)model model.to(device)criterion hinge_loss# 由于初始训练集数量很少所以降低学习率optimizer optim.SGD(model.parameters(), lr1e-4, momentum0.9)# 共训练10轮每隔4论减少一次学习率lr_schduler optim.lr_scheduler.StepLR(optimizer, step_size4, gamma0.1)best_model train_model(data_loaders, model, criterion, optimizer, lr_schduler, num_epochs10, devicedevice)# 保存最好的模型参数save_model(best_model, models/best_linear_svm_alexnet_car.pth)边界框回归训练
准备边界框回归数据集
执行create_bbox_regression_data.py脚本
①读取../../data/voc_car/train/Annotations/中的标注框信息存入bndboxs和../../data/finetune_car/train/Annotations/中的正样本数据存入positive_bndboxes,计算标注框和正样本数据的IoU,针对IoU0.6的正样本数据,保存其到../../data/bbox_regression/positive/中,并保存对应的图片到../../data/bbox_regression/JPEGImages/中,保存对应的标注框信息到../../data/bbox_regression/bndboxs/中,保存对应的图片名到../../data/bbox_regression/car.csv中
以下是create_bbox_regression_data.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/4/3 下午7:19
file: create_bbox_regression_data.py
author: zj
description: 创建边界框回归数据集
import os
import shutil
import numpy as np
import utils.util as util# 正样本边界框数目37222if __name__ __main__:从voc_car/train目录中提取标注边界框坐标从finetune_car/train目录中提取训练集正样本坐标IoU0.5进一步提取IoU0.6的边界框数据集保存在bbox_car目录下voc_car_train_dir ../../data/voc_car/train# ground truthgt_annotation_dir os.path.join(voc_car_train_dir, Annotations)jpeg_dir os.path.join(voc_car_train_dir, JPEGImages)classifier_car_train_dir ../../data/finetune_car/train# positivepositive_annotation_dir os.path.join(classifier_car_train_dir, Annotations)dst_root_dir ../../data/bbox_regression/dst_jpeg_dir os.path.join(dst_root_dir, JPEGImages)dst_bndbox_dir os.path.join(dst_root_dir, bndboxs)dst_positive_dir os.path.join(dst_root_dir, positive)util.check_dir(dst_root_dir)util.check_dir(dst_jpeg_dir)util.check_dir(dst_bndbox_dir)util.check_dir(dst_positive_dir)samples util.parse_car_csv(voc_car_train_dir)res_samples list()total_positive_num 0for sample_name in samples:# 提取正样本边界框坐标IoU0.5positive_annotation_path os.path.join(positive_annotation_dir, sample_name _1.csv)positive_bndboxes np.loadtxt(positive_annotation_path, dtypeint, delimiter )# 提取标注边界框gt_annotation_path os.path.join(gt_annotation_dir, sample_name .xml)bndboxs util.parse_xml(gt_annotation_path)# 计算符合条件IoU0.6的候选建议positive_list list()if len(positive_bndboxes.shape) 1 and len(positive_bndboxes) ! 0:scores util.iou(positive_bndboxes, bndboxs)if np.max(scores) 0.6:positive_list.append(positive_bndboxes)elif len(positive_bndboxes.shape) 2:for positive_bndboxe in positive_bndboxes:scores util.iou(positive_bndboxe, bndboxs)if np.max(scores) 0.6:positive_list.append(positive_bndboxe)else:pass# 如果存在正样本边界框IoU0.6那么保存相应的图片以及标注边界框if len(positive_list) 0:# 保存图片jpeg_path os.path.join(jpeg_dir, sample_name .jpg)dst_jpeg_path os.path.join(dst_jpeg_dir, sample_name .jpg)shutil.copyfile(jpeg_path, dst_jpeg_path)# 保存标注边界框dst_bndbox_path os.path.join(dst_bndbox_dir, sample_name .csv)np.savetxt(dst_bndbox_path, bndboxs, fmt%s, delimiter )# 保存正样本边界框dst_positive_path os.path.join(dst_positive_dir, sample_name .csv)np.savetxt(dst_positive_path, np.array(positive_list), fmt%s, delimiter )total_positive_num len(positive_list)res_samples.append(sample_name)print(save {} done.format(sample_name))else:print(-------- {} 不符合条件.format(sample_name))dst_csv_path os.path.join(dst_root_dir, car.csv)np.savetxt(dst_csv_path, res_samples, fmt%s, delimiter )print(total positive num: {}.format(total_positive_num))print(done)自定义边界框回归训练数据集类
custom_bbox_regression_dataset.py,该脚本不用主动执行,在训练分类器模型的时候,自然会调用到,以下只说这个脚本做了什么事
①BBoxRegressionDataset类继承自Dataset
②__init__时读取../../data/bbox_regression/JPEGImages/文件夹中的图片,存入self.jpeg_list,又读取../../data/bbox_regression/bndboxs/中的标注框信息和../../data/bbox_regression/positive/中的正样本数据并都存入self.box_list
③__getitem__方法计算并返回图片和相对坐标差
以下是custom_bbox_regression_dataset.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/4/3 下午8:07
file: custom_bbox_regression_dataset.py
author: zj
description:
import os
import cv2
import numpy as np
import torch
import torchvision.transforms as transforms
from torch.utils.data import Dataset
from torch.utils.data import DataLoaderimport utils.util as utilclass BBoxRegressionDataset(Dataset):def __init__(self, root_dir, transformNone):super(BBoxRegressionDataset, self).__init__()self.transform transformsamples util.parse_car_csv(root_dir)jpeg_list list()# 保存{image_id: ?, positive: ?, bndbox: ?}box_list list()for i in range(len(samples)):sample_name samples[i]jpeg_path os.path.join(root_dir, JPEGImages, sample_name .jpg)bndbox_path os.path.join(root_dir, bndboxs, sample_name .csv)positive_path os.path.join(root_dir, positive, sample_name .csv)jpeg_list.append(cv2.imread(jpeg_path))bndboxes np.loadtxt(bndbox_path, dtypeint, delimiter )positives np.loadtxt(positive_path, dtypeint, delimiter )if len(positives.shape) 1:bndbox self.get_bndbox(bndboxes, positives)box_list.append({image_id: i, positive: positives, bndbox: bndbox})else:for positive in positives:bndbox self.get_bndbox(bndboxes, positive)box_list.append({image_id: i, positive: positive, bndbox: bndbox})self.jpeg_list jpeg_listself.box_list box_listdef __getitem__(self, index: int):assert index self.__len__(), 数据集大小为%d当前输入下标为%d % (self.__len__(), index)box_dict self.box_list[index]image_id box_dict[image_id]positive box_dict[positive]bndbox box_dict[bndbox]# 获取预测图像jpeg_img self.jpeg_list[image_id]xmin, ymin, xmax, ymax positiveimage jpeg_img[ymin:ymax, xmin:xmax]if self.transform:image self.transform(image)# 计算P/G的x/y/w/htarget dict()p_w xmax - xminp_h ymax - yminp_x xmin p_w / 2p_y ymin p_h / 2xmin, ymin, xmax, ymax bndboxg_w xmax - xming_h ymax - yming_x xmin g_w / 2g_y ymin g_h / 2# 计算tt_x (g_x - p_x) / p_wt_y (g_y - p_y) / p_ht_w np.log(g_w / p_w)t_h np.log(g_h / p_h)return image, np.array((t_x, t_y, t_w, t_h))def __len__(self):return len(self.box_list)def get_bndbox(self, bndboxes, positive):返回和positive的IoU最大的标注边界框:param bndboxes: 大小为[N, 4]或者[4]:param positive: 大小为[4]:return: [4]if len(bndboxes.shape) 1:# 只有一个标注边界框直接返回即可return bndboxeselse:scores util.iou(positive, bndboxes)return bndboxes[np.argmax(scores)]def test():创建数据集类实例transform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.RandomHorizontalFlip(),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])data_root_dir ../../data/bbox_regressiondata_set BBoxRegressionDataset(data_root_dir, transformtransform)print(data_set.__len__())image, target data_set.__getitem__(10)print(image.shape)print(target)print(target.dtype)def test2():测试DataLoader使用transform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.RandomHorizontalFlip(),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])data_root_dir ../../data/bbox_regressiondata_set BBoxRegressionDataset(data_root_dir, transformtransform)data_loader DataLoader(data_set, batch_size128, shuffleTrue, num_workers8)items next(data_loader.__iter__())datas, targets itemsprint(datas.shape)print(targets.shape)print(targets.dtype)if __name__ __main__:test()# test2()训练边界框回归
执行bbox_regression.py脚本
①调用custom_bbox_regression_dataset.py脚本,得到训练数据data_loader
②使用AlexNet网络模型,修改分类器对象classifier的输出为2类(1类是car,一类是背景),加载权重best_linear_svm_alexnet_car.pth,并设置参数冻结,然后再添加一个线性层作为全连接层,定义损失函数为均方误差损失函数,优化函数为SGD,学习率更新器为StepLR,然后开始训练,保存模型到models/bbox_regression_%d.pth
以下是bbox_regression.py脚本代码
# -*- coding: utf-8 -*-
date: 2020/4/3 下午6:55
file: bbox_regression.py
author: zj
description: 边界框回归训练
import os
import copy
import time
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
import torchvision.transforms as transforms
from torchvision.models import AlexNetfrom utils.data.custom_bbox_regression_dataset import BBoxRegressionDataset
import utils.util as utildef load_data(data_root_dir):transform transforms.Compose([transforms.ToPILImage(),transforms.Resize((227, 227)),transforms.RandomHorizontalFlip(),transforms.ToTensor(),transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))])data_set BBoxRegressionDataset(data_root_dir, transformtransform)data_loader DataLoader(data_set, batch_size128, shuffleTrue, num_workers8)return data_loaderdef train_model(data_loader, feature_model, model, criterion, optimizer, lr_scheduler, num_epochs25, deviceNone):since time.time()model.train() # Set model to training modeloss_list list()for epoch in range(num_epochs):print(Epoch {}/{}.format(epoch, num_epochs - 1))print(- * 10)running_loss 0.0# Iterate over data.for inputs, targets in data_loader:inputs inputs.to(device)targets targets.float().to(device)features feature_model.features(inputs)features torch.flatten(features, 1)# zero the parameter gradientsoptimizer.zero_grad()# forwardoutputs model(features)loss criterion(outputs, targets)loss.backward()optimizer.step()# statisticsrunning_loss loss.item() * inputs.size(0)lr_scheduler.step()epoch_loss running_loss / data_loader.dataset.__len__()loss_list.append(epoch_loss)print({} Loss: {:.4f}.format(epoch, epoch_loss))# 每训练一轮就保存util.save_model(model, ./models/bbox_regression_%d.pth % epoch)print()time_elapsed time.time() - sinceprint(Training complete in {:.0f}m {:.0f}s.format(time_elapsed // 60, time_elapsed % 60))return loss_listdef get_model(deviceNone):# 加载CNN模型model AlexNet(num_classes2)model.load_state_dict(torch.load(./models/best_linear_svm_alexnet_car.pth))model.eval()# 取消梯度追踪for param in model.parameters():param.requires_grad Falseif device:model model.to(device)return modelif __name__ __main__:data_loader load_data(./data/bbox_regression)device torch.device(cuda:0 if torch.cuda.is_available() else cpu)feature_model get_model(device)# AlexNet最后一个池化层计算得到256*6*6输出in_features 256 * 6 * 6out_features 4model nn.Linear(in_features, out_features)model.to(device)criterion nn.MSELoss()optimizer optim.Adam(model.parameters(), lr1e-4, weight_decay1e-4)lr_scheduler optim.lr_scheduler.StepLR(optimizer, step_size5, gamma0.1)loss_list train_model(data_loader, feature_model, model, criterion, optimizer, lr_scheduler, devicedevice,num_epochs12)util.plot_loss(loss_list)汽车car目标检测器实现
读取图片,先检测图片中是否有汽车,然后使用非极大值抑制(NMS)算法消除冗余边界框,最后输出目标检测结果,如下图 工程下载
pytorch-r-cnn工程文件
