寮步网站建设极致发烧,h5移动端网站开发,国外云服务器厂商,ui设计是什么专业的目录
1. 数据集处理
1.1 实现脚本
1.2 json文件
2. 设置读取路径
2.1 设置路径
2.2 数据集转换
2.3 数据集预处理
2.4 训练#xff08;3d_fullres)
3. 训练结果展示 关于nnUnet 数据集的处理和环境搭建#xff0c;参考上文#xff1a;第四章#xff1a;nnUnet大模…目录
1. 数据集处理
1.1 实现脚本
1.2 json文件
2. 设置读取路径
2.1 设置路径
2.2 数据集转换
2.3 数据集预处理
2.4 训练3d_fullres)
3. 训练结果展示 关于nnUnet 数据集的处理和环境搭建参考上文第四章nnUnet大模型之环境配置、数据集制作-CSDN博客
1. 数据集处理
因为上文数据集的标签有很多问题虽然处理起来很简单为了防止后续需要这里记录下
观察上文发现数据的标签是19类别但是mask的绘制不是连续的0 1 2 3这样在图像分割中是
不允许的需要做灰度映射。 实际上在做unet一些列多类别分割的时候已经介绍过自适应的灰度映射这里只做简单介绍具体参考下文Unet 实战分割项目、多尺度训练、多类别分割_unet实例分割-CSDN博客 如果数据没有问题的话直接跳到第二章即可 1.1 实现脚本
如下
import SimpleITK as sitk
import numpy as np
import os
from tqdm import tqdm
import shutildef main():root labelsTrimages [os.path.join(root, u) for u in os.listdir(root)]root_ret ret_labelsTrif os.path.exists(root_ret):shutil.rmtree(root_ret)os.mkdir(root_ret)# 计算灰度cl []for i in tqdm(images, descprocess):mask sitk.ReadImage(i)mask sitk.GetArrayFromImage(mask)mask np.unique(mask)for h in mask:if h not in cl:cl.append(h)cl.sort()n len(cl)print(cl) # [0, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22]print(分割的个数:,n)if n cl[n-1]:return# 灰度映射for i in tqdm(images, descprocess):mask sitk.ReadImage(i)mask sitk.GetArrayFromImage(mask)for index,h in enumerate(cl):mask[maskh] indexmask sitk.GetImageFromArray(mask)ret_path i.replace(root,root_ret)sitk.WriteImage(mask,ret_path)# 检查灰度cl_ret []images [os.path.join(root_ret, u) for u in os.listdir(root_ret)]for i in tqdm(images, descprocess):mask sitk.ReadImage(i)mask sitk.GetArrayFromImage(mask)mask np.unique(mask)for h in mask:if h not in cl_ret:cl_ret.append(h)cl_ret.sort()n len(cl_ret)print(cl_ret) # [0, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22]print(处理后分割的个数:,n)if __name__ __main__:main()摆放如下脚本会将labelsTr的标签自动映射成0 1 2 3连续的并且保存在新生成的ret下 运行如下 可以看到mask的灰度已经进行了映射 通过itk打开可以发现mask并没有改变只是里面的数字变了这样颜色显示也就变了
源标签 处理完的 1.2 json文件
更改如下当然新的json文件可以用上文的脚本生成
{labels: {0: background,1: L1,2: L2,3: L3,4: L4,5: L5,6: L6,7: L7,8: L8,9: L9,10: L10,11: L11,12: L12,13: L13,14: L14,15: L15,16: L16,17: L17,18: L18},modality: {0: CT},numTest: 0,numTraining: 40,tensorImageSize: 3D,test: [],training: [{image: ./imagesTr/spine_001.nii.gz,label: ./labelsTr/spine_001.nii.gz},{image: ./imagesTr/spine_002.nii.gz,label: ./labelsTr/spine_002.nii.gz},{image: ./imagesTr/spine_003.nii.gz,label: ./labelsTr/spine_003.nii.gz},{image: ./imagesTr/spine_004.nii.gz,label: ./labelsTr/spine_004.nii.gz},{image: ./imagesTr/spine_005.nii.gz,label: ./labelsTr/spine_005.nii.gz},{image: ./imagesTr/spine_006.nii.gz,label: ./labelsTr/spine_006.nii.gz},{image: ./imagesTr/spine_007.nii.gz,label: ./labelsTr/spine_007.nii.gz},{image: ./imagesTr/spine_008.nii.gz,label: ./labelsTr/spine_008.nii.gz},{image: ./imagesTr/spine_009.nii.gz,label: ./labelsTr/spine_009.nii.gz},{image: ./imagesTr/spine_010.nii.gz,label: ./labelsTr/spine_010.nii.gz},{image: ./imagesTr/spine_011.nii.gz,label: ./labelsTr/spine_011.nii.gz},{image: ./imagesTr/spine_012.nii.gz,label: ./labelsTr/spine_012.nii.gz},{image: ./imagesTr/spine_013.nii.gz,label: ./labelsTr/spine_013.nii.gz},{image: ./imagesTr/spine_014.nii.gz,label: ./labelsTr/spine_014.nii.gz},{image: ./imagesTr/spine_015.nii.gz,label: ./labelsTr/spine_015.nii.gz},{image: ./imagesTr/spine_016.nii.gz,label: ./labelsTr/spine_016.nii.gz},{image: ./imagesTr/spine_017.nii.gz,label: ./labelsTr/spine_017.nii.gz},{image: ./imagesTr/spine_018.nii.gz,label: ./labelsTr/spine_018.nii.gz},{image: ./imagesTr/spine_019.nii.gz,label: ./labelsTr/spine_019.nii.gz},{image: ./imagesTr/spine_020.nii.gz,label: ./labelsTr/spine_020.nii.gz},{image: ./imagesTr/spine_021.nii.gz,label: ./labelsTr/spine_021.nii.gz},{image: ./imagesTr/spine_022.nii.gz,label: ./labelsTr/spine_022.nii.gz},{image: ./imagesTr/spine_023.nii.gz,label: ./labelsTr/spine_023.nii.gz},{image: ./imagesTr/spine_024.nii.gz,label: ./labelsTr/spine_024.nii.gz},{image: ./imagesTr/spine_025.nii.gz,label: ./labelsTr/spine_025.nii.gz},{image: ./imagesTr/spine_026.nii.gz,label: ./labelsTr/spine_026.nii.gz},{image: ./imagesTr/spine_027.nii.gz,label: ./labelsTr/spine_027.nii.gz},{image: ./imagesTr/spine_028.nii.gz,label: ./labelsTr/spine_028.nii.gz},{image: ./imagesTr/spine_029.nii.gz,label: ./labelsTr/spine_029.nii.gz},{image: ./imagesTr/spine_030.nii.gz,label: ./labelsTr/spine_030.nii.gz},{image: ./imagesTr/spine_031.nii.gz,label: ./labelsTr/spine_031.nii.gz},{image: ./imagesTr/spine_032.nii.gz,label: ./labelsTr/spine_032.nii.gz},{image: ./imagesTr/spine_033.nii.gz,label: ./labelsTr/spine_033.nii.gz},{image: ./imagesTr/spine_034.nii.gz,label: ./labelsTr/spine_034.nii.gz},{image: ./imagesTr/spine_035.nii.gz,label: ./labelsTr/spine_035.nii.gz},{image: ./imagesTr/spine_036.nii.gz,label: ./labelsTr/spine_036.nii.gz},{image: ./imagesTr/spine_037.nii.gz,label: ./labelsTr/spine_037.nii.gz},{image: ./imagesTr/spine_038.nii.gz,label: ./labelsTr/spine_038.nii.gz},{image: ./imagesTr/spine_039.nii.gz,label: ./labelsTr/spine_039.nii.gz},{image: ./imagesTr/spine_040.nii.gz,label: ./labelsTr/spine_040.nii.gz}]
} 2. 设置读取路径
回到正文这里的Task下有如下数据source nnunet/bin/activate 激活nnunet环境 Tips这里的 labelsTr和dataset.json是第一节处理后的 任务名称为Task01_Spine 2.1 设置路径
这里设置为绝对路径除了DATASET后面的前面部分需要根据不同机器设定
在这里更改 vim .bashrcvim ~/.bashrc 末尾最后面
export nnUNet_raw_data_base/*/DATASET/nnUNet_raw
export nnUNet_preprocessed/*/DATASET/nnUNet_preprocessed
export RESULTS_FOLDER/*/DATASET/nnUNet_trained_models这里设置后如果想要训练其他模型不需要在进行更改
添加完成后保存 source ~/.bashrc 更新环境变量可以通过echo $RESULTS_FOLDER 检查是否修改成功 2.2 数据集转换
下面命令都是在environments 目录里进行操作 转换命令为
nnUNet_convert_decathlon_task -i DATASET/nnUNet_raw/nnUNet_raw_data/Task01_Spine/转换完的数据在 图像可能具有多种模态nnU-Net通过其后缀文件名末尾的四位整数识别成像模态。因此图像文件必须遵循以下命名约定case_identifier_XXXX.nii.gz。 这里XXXX是模态标识符。dataset.json文件中指定了这些标识符所属的模态。 标签文件保存为case_identifier.nii.gz 例如BrainTumor。每个图像有四种模态FLAIR0000、T1w0001、T1gd0002和T2w0003 2.3 数据集预处理
命令如下这里只会做训练集进行预处理测试集不会处理
nnUNet_plan_and_preprocess -t 1
只需要一行命令因为 Task_id是1所以这里的数字就是1。这个过程会消耗很多的时间速度慢的原因在于对要进行插值等各种操作。 生成的数据在crop和precocessed里面查看 2.4 训练3d_fullres)
命令如下
nnUNet_train 3d_fullres nnUNetTrainerV2 1 01 指的是Task标号5 指定训练的是5倍交叉验证的哪一倍。 会实时生成如下结果在这里 nnUNet_trained_models 3. 训练结果展示
RTX 3090跑一个epoch大概100s1000个epoch估计要一两天等跑完下篇文章在贴训练结果吧
