前面学习了一下论文：Improved Field-Based Soybean Seed Counting and Localization with Feature Level Considered

论文链接：https://spj.science.org/doi/10.34133/plantphenomics.0026

解读链接：论文阅读--考虑特征水平的改进的基于田间的大豆种子计数和定位_追忆苔上雪的博客-CSDN博客

下面开始着手复现p2p大豆计数模型

(一)p2p人群计数源码以及复现

由于p2p大豆计数模型建立在p2p人群计数，这里提供一下p2p人群计数代码

<1>p2p人群计数源码：https://github.com/TencentYoutuResearch/CrowdCounting-P2PNet

<2>p2p人群计数源码复现过程：

crowdcountingp2p代码复现_追忆苔上雪的博客-CSDN博客

crowdcountingp2p代码复现(续)_追忆苔上雪的博客-CSDN博客

(二)p2p大豆计数代码复现

<1>p2p大豆计数源码

源码地址：https://github.com/UTokyo-FieldPhenomics-Lab/P2PNet-Soy

由于该开源代码是用jupyter写的，不能利用pycharm和pytorch直接进行复现，下面开始介绍利用自己pycharm复现的方法

<2>植物计数方法的拓展

在复现代码之前，先介绍一下植物计数过程需要使用到的方法，大家按需自取，链接附上：

作物计数方法汇总_追忆苔上雪的博客-CSDN博客

作物计数方法之合并信息生成json标签的方法_追忆苔上雪的博客-CSDN博客

这里再拓展一个将标签中提到搭配点的列表完整存入txt文本的方法，方便将p2p大豆计数方法拓展至其他植物

import scipy.io as sio
import numpy as np
import os
mat_filename_list = os.listdir(r"D:\P2P_plant_counting\P2P_watermelon\watermelon_dataset\data_root\test_data\label_mat")
mat_filename_list2 = os.listdir(r"D:\P2P_plant_counting\P2P_watermelon\watermelon_dataset\data_root\train_data\label_mat")
# print(mat_filename_list)
# 批量转化 mat 文件为 txt 文件并保存在 txt_file 文件夹中
for mat_filename in mat_filename_list:# print(mat_filename)matdata = sio.loadmat("D:/P2P_plant_counting/P2P_watermelon/watermelon_dataset/data_root/test_data/label_mat/" + mat_filename)# print(matdata)data = matdata["image_info"]array1 = data[0][0][0][0][0]# print(array1)for i in array1:# print(i)a1 = i[0]b1 = i[1]# print(list1)txt_filename = mat_filename.split('.')[0]# print(txt_filename)with open(r'D:\P2P_plant_counting\P2P_watermelon\watermelon_dataset\data_root\test_data\label\{}.txt'.format(txt_filename), 'a') as fp:  # a 防止覆盖循环写入fp.write('['+str(a1)+','+str(b1)+']'+'\n')fp.close()
for mat_filename2 in mat_filename_list2:matdata2 = sio.loadmat("D:/P2P_plant_counting/P2P_watermelon/watermelon_dataset/data_root/train_data/label_mat/" + mat_filename2)# print(matdata)data2 = matdata2["image_info"]array2 = data2[0][0][0][0][0]# print(array1)for j in array1:# print(i)a2 = j[0]b2 = j[1]# print(list1)txt_filename2 = mat_filename2.split('.')[0]# print(txt_filename)with open(r'D:\P2P_plant_counting\P2P_watermelon\watermelon_dataset\data_root\train_data\label\{}.txt'.format(txt_filename2), 'a') as fp:  # a 防止覆盖循环写入fp.write('['+str(a2)+','+str(b2)+']'+'\n')fp.close()

则可以将mat标签中提取到点的信息存到txt文本中方便后续训练所需

<3>p2p大豆计数框架

p2p大豆计数框架如下图所示，其中数据集在源码中的链接中有下载链接，这里不多做介绍

若是需要自己将p2p计数框架用于其作物，数据制作可以参考上文链接中所介绍的方法

vgg16的权重文件也可自己下载：这里提供链接:

https://download.pytorch.org/models/vgg16_bn-6c64b313.pth

https://download.pytorch.org/models/vgg16-397923af.pth

<4>关键代码介绍

其中数据处理代码相比源码，这里做了一点改变，详见下图

训练代码参数解释

parser.add_argument('--point_loss_coef', default=0.0002, type=float)  # default = 0.0002 # 0.5
# the final classification loss = -(the sum of positive confidence score + eos_coef * the sum of negative confidence score)/M, M is total proposed points
parser.add_argument('--eos_coef', default=0.05, type=float,  # 0.05help="Relative classification weight of the no-object class")  # default = 0.5# a threshold during evaluation for counting and visualization
parser.add_argument('--threshold', default=0.5, type=float,help="threshold in evalluation: evaluate_crowd_no_overlap")
parser.add_argument('--row', default=2, type=int,help="row number of anchor points")
parser.add_argument('--line', default=2, type=int,help="line number of anchor points")# dataset parameters
parser.add_argument('--dataset_file', default='SOYBEAN')
parser.add_argument('--data_root', default='/home/king/Projects/LiuHuaiyang/p2p_plant_counting/p2p_Soybean/soybean_datasets/SOYBEAN/data_root',help='path where the dataset is')parser.add_argument('--output_dir', default='output_dir',   # 存放输出日志help='path where to save, empty for no saving')
parser.add_argument('--checkpoints_dir',default='checkpointds_dir',       # 存放权重路径help='path where to save checkpoints, empty for no saving')    # ckpt_5n was not bad, default 2 X 2
parser.add_argument('--tensorboard_dir',default='tensorboard_dir',       # 存放训练阶段和评估阶段loss值help='path where to save, empty for no saving')

parser.add_argument('--data_root用于存放数据集路径

parser.add_argument('--output_dir用于存放输出日志

parser.add_argument('--checkpoints_dir',用于存放训练的权重路径

parser.add_argument('--tensorboard_dir',用于存放训练阶段和评估阶段loss值

推理代码参数解释

parser = argparse.ArgumentParser(description="Object Counting Framework")
# * Backbone
parser.add_argument('--backbone', default='vgg16', type=str,help="Name of the convolutional backbone to use")
#vgg16_bn
# a threshold during evaluation for counting and visualization
parser.add_argument('--threshold', default=0.5, type=float,help="threshold in evalluation: evaluate_crowd_no_overlap")
parser.add_argument('--row', default=2, type=int,help="row number of anchor points")
parser.add_argument('--line', default=2, type=int,help="line number of anchor points")
parser.add_argument('--data_root', default='/home/king/Projects/LiuHuaiyang/p2p_plant_counting/p2p_Soybean/soybean_datasets/SOYBEAN/data_root/',help='path where the dataset is')
parser.add_argument('--seed', default=42, type=int)
parser.add_argument('--resume', default="/home/king/Projects/LiuHuaiyang/p2p_plant_counting/p2p_Soybean/checkpointds_dir/best_mae.pth", help='resume from checkpoint')  # 训练好的权重位置
parser.add_argument('--vis_dir', default='/media/king/DATA/LiuHuaiyang/p2p_plant_counting/p2p_Soybean/vis_p2pnetSoy_out')    # 预测输出路径
parser.add_argument('--eval', action='store_true')
parser.add_argument('--num_workers', default=1, type=int)
parser.add_argument('--gpu_id', default=0, type=int, help='the gpu used for training')
parser.add_argument('--half', default=True, type=bool)

parser.add_argument('--resume'用于存放训练好的权重位置

parser.add_argument('--vis_dir'为预测输出路径