deeplabcut使用

cuda11.2和cudnn8.1安装

1. win
   最新的 Win11/WIN10 安装CUDA11.2和cuDNN
2. linux
   ubuntu 16.04 安装 cuda11.2 和cudnn8.2.1

dlc安装

1. 创建虚拟环境
   
   

2. 安装deeplabcut=2.2.3 tensorflow=2.11.0 wxPython=4.0.4
   pip install deeplabcut==2.2.3
   deeplabcut里面包含了tensorflow的不用再安装了

pip install tensorflow-gpu==2.12.0 -i https://pypi.tuna.tsinghua.edu.cn/simple/
pip install deeplabcut==2.2.3  -i https://pypi.tuna.tsinghua.edu.cn/simple/
pip install wxPython==4.0.4  -i https://pypi.tuna.tsinghua.edu.cn/simple/
-c https://mirrors.tuna.tsinghua.edu.cn/anaconda/pkgs/main/
2023-03-08 22:26:34.816159: W tensorflow/stream_executor/platform/default/dso_loader.cc:64] Could not load dynamic library 'cudart64_110.dll'; dlerror: cudart64_110.dll not found
2023-03-08 22:26:34.816852: I tensorflow/stream_executor/cuda/cudart_stub.cc:29] Ignore above cudart dlerror if you do not have a GPU set up on your machine.

基本框架

本地操作（本地标注，服务器训练。gpu太low了）

官网教程

激活环境并进入ipython

linux系统有的用ipython不行就用python

导入deeplabcut

import deeplabcut 

创建文件

deeplabcut.create_new_project(r'project-name',r'Laborator',working_dirctory=[r'vidoes_path'],copy_videos=True)

配置文件

config.yaml

1. bodyparts:
2. numframes2pick
3. skeleton:==注意字母不能写错了

config_path = 'G:\\new\\dlc\\dlc_mouse_4-cunyiyuan-2022-10-27\\config.yaml'

抽帧

deeplabcut.extract_frames(config)

错误

原因：配置文件中不能有中文

打标签

deeplabcut.label_frames(config_path)

检查标签

deeplabcut.check_labels(config_path)

服务器训练操作

import deeplabcut
config_path=‘’

Create a training dataset

deeplabcut.create_training_dataset(config_path, net_type='resnet_50', augmenter_type='imgaug')

Start training

deeplabcut.train_network(config_path, shuffle=1, trainingsetindex=0, gputouse=None, max_snapshots_to_keep=5, autotune=False, displayiters=100, saveiters=15000, maxiters=30000, allow_growth=True)

Evaluate the Trained Network

deeplabcut.evaluate_network(config_path,Shuffles=[1], plotting=True)

Video analysis

deeplabcut.analyze_videos(config_path, ['fullpath/analysis/project/videos'], videotype='avi', shuffle=1, trainingsetindex=0, gputouse='0', save_as_csv=True, destfolder=None, dynamic=(True, .5, 10))

filter

deeplabcut.filterpredictions(config_path,['fullpath/analysis/project/videos'], videotype='.mp4',filtertype= 'arima',ARdegree=5,MAdegree=2)

Plot Trajectories

deeplabcut.plot_trajectories(config_path, [‘fullpath/analysis/project/videos/reachingvideo1.avi’])

Create Labeled Videos

deeplabcut.create_labeled_video(config_path, videofile_path, videotype=VideoType, draw_skeleton=True, save_frames=True, trailpoints=0)

Extract “Skeleton” Features

deeplabcut.analyzeskeleton(config, video, videotype='avi', shuffle=1, trainingsetindex=0, save_as_csv=False, destfolder=None)

outlier

deeplabcut.extract_outlier_frames(config_path, ['videofile_path'], outlieralgorithm='manual')

refine

deeplabcut.refine_labels(config_path)
deeplabcut.merge_datasets(config_path)

恢复训练

deeplabcut源代码：

• 训练错误
  (0) DATA_LOSS: Unable to open table file /home/study11/szc/hw_01-szc-2023-03-07/dlc-models/iteration-0/hw_01Mar7-trainset80shuffle1/train/snapshot-100000.data-00000-of-00001: DATA_LOSS: not an sstable (bad magic number): perhaps your file is in a differ

#解决

四个权重文件，分别为：
1.checkpoint
2.snapshot-100.index
3.snapshot-100.meta
4.snapshot-100.data-00000-of-00001

那么导入时，init_weights应填"snapshot-100"的路径，
不是“snapshot-100.data-00000-of-00001”，也不是“snapshot”！
deeplabcut.train_network(config_path, shuffle=1, trainingsetindex=0, gputouse='0,1,2', max_snapshots_to_keep=10, autotune=False, displayiters=100, saveiters=10000, maxiters=200000, allow_growth=True)

3D操作

相机标定

• 采集成对图片
  用get_paired_image.py采集标定的照片，复制到DLC创建的工程文件–标定文件夹里D:\Liqiqi\DeepLabCut_Liqiqi\liqiqi01-mouse-2022-12-12-3d\calibration_images】
• 摄像头标定
  【（二）创建3D工程文件，注意：创建好新的工程文件之后，在标定的时候，config文件不需要改参数】
  工程文件创建成功后包括：calibration_images, camera_matrix, corners, and undistortion五个文件夹
• calibration_images：该目录将包含从两台相机获取的一组校准图像。拍摄图像时：注意命名camera-1-01.jpg或者camera-2-01.jpg
• camera_matrix：此目录将两台摄像机的参数存储为泡菜文件
• corners：作为相机校准的一部分，在校准图像中检测到棋盘图案，这些图案将存储在此目录中
• undistortion：为了检查校准，校准图像和相应的角点不会失真。这些未失真的图像被未失真的点覆盖，并将存储在此目录中。
• 标定
  deeplabcut.calibrate_cameras(config_path3d, cbrow =8,cbcol =11,calibrate=False,alpha=0.9) # False是移除不好的照片
  deeplabcut.calibrate_cameras(config_path3d, cbrow =8,cbcol =11,calibrate=True,alpha=0.9) # Ture 是对好的照片进行标定

3d使用

官网教程

1. create project

deeplabcut.create_new_project_3d('ProjectName','NameofLabeler',working_directory='path',num_cameras = 2)

2. config_path3d

config_path3d = r'/content/drive/MyDrive/Colab Notebooks/39d/config.yaml'

deeplabcut.calibrate_cameras(config_path3d, cbrow =8,cbcol =11,calibrate=False,alpha=0.9) 

deeplabcut.calibrate_cameras(config_path3d, cbrow =8,cbcol =11,calibrate=True,alpha=0.9)

deeplabcut.check_undistortion(config_path3d, cbrow=8, cbcol=11)

deeplabcut.triangulate(config_path3d, r'videopath', filterpredictions=True, videotype='avi', save_as_csv=True, gputouse='0')

deeplabcut.create_labeled_video_3d(config_path3d, [r'/content/drive/MyDrive/Colab Notebooks/hw_01-szc-2023-03-07/videos'], start=200, end=600,videotype='mp4', xlim =[-50,50], ylim=[-30,30], zlim=[-30,30],view=[45,270])