Skip to content

Latest commit

 

History

History
38 lines (21 loc) · 1.79 KB

Quicktraining-guide.md

File metadata and controls

38 lines (21 loc) · 1.79 KB
Raw

Quick Training Guide:

After becoming familiar with the workflow, here is the list of commands to run in the terminal for training. For more information on the output of these programs etc., check out the detailed walk-through with labeled example data.

You will need to adjust the myconfig.py file first (and you will need to adjust your folder names below accordingly):

(A) Place your videos in the folder DeepLabCut-master/Generating_a_Training_Set, first select and label frames:

  $ python3 Step1_SelectRandomFrames_fromVideos.py

(B) Then (after labeling all subfolders) create a pandas array with the data:

  $ python3 Step2_ConvertingLabels2DataFrame.py

(C) Check labels, and prepare to run:

  $ python3 Step3_CheckLabels.py  #go check the frames in the newly created folders, then if okay, run:
  $ python3 Step4_GenerateTrainingFileFromLabelledData.py

(D) Transfer the folders just created:

   $ cp -R YOURexperimentNameTheDate-trainset95shuffle1 ../pose-tensorflow/models/

   $ cp -R UnaugmentedDataSet_YOURexperimentNameTheDate/ ../pose-tensorflow/models/

(E) Download the pre-trained weights (if not already done):

  $ cd ../pose-tensorflow/models/pretrained
  $ ./download.sh

(F) Train:

  $ cd ../YOURexperimentNameTheDate-trainset95shuffle1/train #change this to your folder name!
  $ TF_CUDNN_USE_AUTOTUNE=0 CUDA_VISIBLE_DEVICES=0 python3 ../../../train.py

Once your network generalizes well as tested by evaluation tools, go to the analysis tools to extract poses from videos. For more information on the output of these programs etc., check out the detailed walk-through with labeled example data.

Return to readme.