Skip to content

Latest commit

 

History

History
317 lines (243 loc) · 12.3 KB

BehavioralClustering.md

File metadata and controls

317 lines (243 loc) · 12.3 KB
Raw

Behavioral Clustering

Algorithm overview

The main process of hierarchical clustering list below can be found in ./fft_main_sep_twoMiceInteract.py

  1. Get data from tracking result with function retrieve_poses_Mice
  2. Clean the data with functions clean_differentLength_clips, remove_longMiceDist_clips, left_right, correctLimbs
  3. Preprocess the data with function align_poses_self
  4. Get feature from the data with funcion compute_features_sep_twoMice_Independent
  5. Conduct hierarchical clustering with function cluster
  6. Visualize and save the result
  7. Generate input files for clustering UI

Run By GUI (recommended for non-cs users)



AlphaTracker GUI behavior clustering page

Please visit our video tutorial for behavior clustering at YouTube.

Or run by command line

Step 1. Configuration

Set the Behavioral Clustering folder as the current directory.

cd BehavioralClustering

Change the following settings in ./setting.py. Changes for individual behavior and social behavior are explained separately here.

Configuration for individual behavior

  • Adjust the weights for Individual behavioral features
        # cluster_arg: parameter of features and threshold for clustering
        self.cluster_arg = [
        {
            'thred':30, # threshold for defining cluster in the dendrogram
            'name':'all_twoMice', ## name: for display
            'evaluation_metric':'Adjusted Rand index',
            'features_arg':[\
             # {'feat_key':'newFeatureName','weight':4,'norm':'zscore_all'},
             {'feat_key':'left_ear','weight':1,'norm':'zscore_all'},
             {'feat_key':'right_ear','weight':1,'norm':'zscore_all'},
             {'feat_key':'left_ear_phi','weight':1,'norm':'zscore_all'},
             {'feat_key':'right_ear_phi','weight':1,'norm':'zscore_all'},
             {'feat_key':'body_length','weight':1,'norm':'zscore_all'},
             {'feat_key':'head_length','weight':1,'norm':'zscore_all'},
             {'feat_key':'head_body_angles','weight':1,'norm':'zscore_all'},
             {'feat_key':'displace_rho','weight':1,'norm':'zscore_all'},
             {'feat_key':'displace_phi_c','weight':1,'norm':'zscore_all'},
             {'feat_key':'displace_phi_s','weight':1,'norm':'zscore_all'},
             {'feat_key':'nose_fft_amp','weight':1,'norm':'zscore_all'},
             {'feat_key':'nose_fft_ang','weight':1,'norm':'zscore_all'},
             {'feat_key':'contourPCA_fft_amp','weight':1,'norm':'zscore_all'},
             {'feat_key':'contourPCA_fft_ang','weight':1,'norm':'zscore_all'},
             {'feat_key':'body_change_ang','weight':1,'norm':'zscore_all'},

             {'feat_key':'left_ear_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'right_ear_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'left_ear_phi_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'right_ear_phi_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'body_length_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'head_length_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'head_body_angles_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'displace_rho_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'displace_phi_c_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'displace_phi_s_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'nose_fft_amp_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'nose_fft_ang_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'contourPCA_fft_amp_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'contourPCA_fft_ang_TO','weight':0,'norm':'zscore_all'},
             {'feat_key':'body_change_ang_TO','weight':0,'norm':'zscore_all'},

             {'feat_key':'TM_nose_RM_tail_displace_phi','weight':0,'norm':'zscore_all'},
             {'feat_key':'TM_nose_RM_tail_displace_rho','weight':0,'norm':'zscore_all'},
             {'feat_key':'RM_nose_TM_tail_displace_phi','weight':0,'norm':'zscore_all'},
             {'feat_key':'RM_nose_TM_tail_displace_rho','weight':0,'norm':'zscore_all'},
             {'feat_key':'nose_nose_displace_rho','weight':0,'norm':'zscore_all'},
             {'feat_key':'nose_nose_displace_phi','weight':0,'norm':'zscore_all'},
             {'feat_key':'two_body_ang','weight':0,'norm':'zscore_all'},
             {'feat_key':'two_head_ang','weight':0,'norm':'zscore_all'},
             ],
          },
        ]
  • Change the threshold for distance between two mice to a very large number
self.distance_threshold = 10000 #
  • Specify the directory to save the generated videos
self.gen_video_folder = # where the generated videos will be saved
  • Specify the directory to save the metadata which will be used for inspecting the clustering results
self.result_folder = # where the metadata will be saved
  • Specify the suffix for generated videos
self.video_name_suffix = # suffix for generated videos

Configurations for social behavior

  • Adjust the weights for Social behavioral features

    self.cluster_arg = [
      {
          'thred':30, # threshold for defining cluster in the dendrogram
          'name':'all_twoMice',## name: for display
          'evaluation_metric':'Adjusted Rand index',
          'features_arg':[\
           # {'feat_key':'newFeatureName','weight':4,'norm':'zscore_all'},
           {'feat_key':'left_ear','weight':1,'norm':'zscore_all'},
           {'feat_key':'right_ear','weight':1,'norm':'zscore_all'},
           {'feat_key':'left_ear_phi','weight':1,'norm':'zscore_all'},
           {'feat_key':'right_ear_phi','weight':1,'norm':'zscore_all'},
           {'feat_key':'body_length','weight':1,'norm':'zscore_all'},
           {'feat_key':'head_length','weight':1,'norm':'zscore_all'},
           {'feat_key':'head_body_angles','weight':1,'norm':'zscore_all'},
           {'feat_key':'displace_rho','weight':1,'norm':'zscore_all'},
           {'feat_key':'displace_phi_c','weight':1,'norm':'zscore_all'},
           {'feat_key':'displace_phi_s','weight':1,'norm':'zscore_all'},
           {'feat_key':'nose_fft_amp','weight':1,'norm':'zscore_all'},
           {'feat_key':'nose_fft_ang','weight':1,'norm':'zscore_all'},
           {'feat_key':'contourPCA_fft_amp','weight':1,'norm':'zscore_all'},
           {'feat_key':'contourPCA_fft_ang','weight':1,'norm':'zscore_all'},
           {'feat_key':'body_change_ang','weight':1,'norm':'zscore_all'},

           {'feat_key':'left_ear_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'right_ear_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'left_ear_phi_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'right_ear_phi_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'body_length_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'head_length_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'head_body_angles_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'displace_rho_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'displace_phi_c_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'displace_phi_s_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'nose_fft_amp_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'nose_fft_ang_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'contourPCA_fft_amp_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'contourPCA_fft_ang_TO','weight':1,'norm':'zscore_all'},
           {'feat_key':'body_change_ang_TO','weight':1,'norm':'zscore_all'},

           {'feat_key':'TM_nose_RM_tail_displace_phi','weight':4,'norm':'zscore_all'},
           {'feat_key':'TM_nose_RM_tail_displace_rho','weight':4,'norm':'zscore_all'},
           {'feat_key':'RM_nose_TM_tail_displace_phi','weight':4,'norm':'zscore_all'},
           {'feat_key':'RM_nose_TM_tail_displace_rho','weight':4,'norm':'zscore_all'},
           {'feat_key':'nose_nose_displace_rho','weight':4,'norm':'zscore_all'},
           {'feat_key':'nose_nose_displace_phi','weight':4,'norm':'zscore_all'},
           {'feat_key':'two_body_ang','weight':4,'norm':'zscore_all'},
           {'feat_key':'two_head_ang','weight':4,'norm':'zscore_all'},
           ],
        },
      ]

  • Change the threshold for distance between two mice to include only mice that are within 1.8 body length

self.distance_threshold = 1.8
  • Specify the directory to save the generated videos
self.gen_video_folder = # where the generated videos will be saved
  • Specify the directory to save the metadata which will be used for inspecting the clustering results
self.result_folder = # where the metadata will be saved
  • Specify the suffix for generated videos
self.video_name_suffix = # suffix for generated videos

step 2. Preprocess Data

Run the following command in the current folder

bash run_all.sh

step 3. Clustering

Run the following command in the current folder

python fft_main_sep_twoMiceInteract.py

Add new features (optional)

There are two steps to add feature for clustering:

step 1. Add Code for the Feature

  1. Define the new feature ./fft_utils.py as the following template (you can find this template in ./fft_utils.py ；
        ## you can process data to get the feature
        ## the pose data are in five variables: pose_clips, pose_clips_align, poseTheOther_clips, poseTheOther_clips_alignSelf, poseTheOther_clips_alignToOther
        ## each of the variables is a numpy array whose shape is (number_of_clip, number_of_frames_in_one_clip, number_of_key_point, 3)
        ## pose_clips contains the raw key points data of the target mouse
        ## pose_clips_align contains key point data that is aligned to the target mouse in the middle frame
        ## poseTheOther_clips contains the raw key points data of the mouse that is closest to the target mouse
        ## poseTheOther_clips_alignSelf contains the key points data of the closest mousethat that is aligned to itself in the middle frame
        ## poseTheOther_clips_alignToOther contains the key points data of the closest mousethat that is aligned to the target mouse in the middle frame
        newFeature = np.ones(pose_clips.shape[1])  # the feature of one clip should be a numpy array whose shape is the frame number of the clip
        if 'newFeatureName' in feature_clips_dict:
            feature_clips_dict['newFeatureName'].append(newFeature)
        else:
            feature_clips_dict['newFeatureName'] = [newFeature]
  1. define the weight and normalization of the new feature in ./utils_file/setting.py :
	self.cluster_arg = [
           {
            'thred':30,
            'name':'all_twoMice',
            'evaluation_metric':'Adjusted Rand index',
            'features_arg':[\
             {'feat_key':'newFeatureName','weight':4,'norm':'zscore_all'}, # add the setting of the new feature here

             {'feat_key':'body_length','weight':1,'norm':'zscore_all'},
             {'feat_key':'head_length','weight':1,'norm':'zscore_all'},
             {'feat_key':'head_body_angles','weight':1,'norm':'zscore_all'},
             {'feat_key':'displace_rho','weight':1,'norm':'zscore_all'},
             {'feat_key':'displace_phi_c','weight':1,'norm':'zscore_all'},
             ],
             },
        ]

step 2. Clustering

Run the following command in the current folder

python fft_main_sep_twoMiceInteract.py

Inspect Clustering results


Run Analysis.ipynb to generate the following plots

Detailed instructions are included in the jupyter notebook

  • Dendrogram

  • Timeline

  • UMAP

  • Mutual information plots

  • Similarity matrix between clusters


Run Visualize.ipynb to visualize skeletons for each cluster

Detailed instructions are included in the jupyter notebook

  • Individual cluster skeletons

  • Social behaviror cluster skeletons