Behavior transition probabilities¶

This script calculates the transition probabilities for zebra behaviors (Grevy's and Plains) from the subset of KABR sessions originally included in the initial mini-scene release.

Download dataset from HuggingFace¶

Use 'consolidated_metadata.csv' file from kabr-datapalooza-2023 for data used in this case study.

In [2]:

import pandas as pd
import seaborn as sns
import matplotlib.pyplot as plt
import warnings
warnings.filterwarnings('ignore')

import pandas as pd import seaborn as sns import matplotlib.pyplot as plt import warnings warnings.filterwarnings('ignore')

In [3]:

data = pd.read_csv('data/consolidated_metadata.csv')

data = pd.read_csv('data/consolidated_metadata.csv')

In [4]:

zebras = data[data['label'] == 'Zebra']
zebras = zebras[['frame', 'id', 'behaviour', 'video']]
zebras['time'] = pd.to_datetime(zebras['frame'], unit='s') - pd.to_datetime(zebras['frame'], unit='s').min()

zebras = data[data['label'] == 'Zebra'] zebras = zebras[['frame', 'id', 'behaviour', 'video']] zebras['time'] = pd.to_datetime(zebras['frame'], unit='s') - pd.to_datetime(zebras['frame'], unit='s').min()

In [5]:

zebras['behaviour'].unique()

zebras['behaviour'].unique()

Out[5]:

array(['Walk', 'Head Up', 'Out of Frame', 'Graze', 'Occluded',
       'Out of Focus', 'Auto-Groom', 'Defecating', 'Trotting', 'Running',
       'Sniff', 'Mutual Grooming', 'Urinating', 'Fighting'], dtype=object)

In [7]:

zebras.head(12)

zebras.head(12)

Out[7]:

	frame	id	behaviour	video	time
0	0	2.0	Walk	12_01_23-DJI_0994	0 days 00:00:00
1	0	1.0	Walk	12_01_23-DJI_0994	0 days 00:00:00
2	1	2.0	Walk	12_01_23-DJI_0994	0 days 00:00:01
3	1	1.0	Walk	12_01_23-DJI_0994	0 days 00:00:01
4	2	2.0	Walk	12_01_23-DJI_0994	0 days 00:00:02
5	2	1.0	Walk	12_01_23-DJI_0994	0 days 00:00:02
6	3	2.0	Walk	12_01_23-DJI_0994	0 days 00:00:03
7	3	1.0	Walk	12_01_23-DJI_0994	0 days 00:00:03
8	4	2.0	Walk	12_01_23-DJI_0994	0 days 00:00:04
9	4	1.0	Walk	12_01_23-DJI_0994	0 days 00:00:04
10	5	1.0	Walk	12_01_23-DJI_0994	0 days 00:00:05
11	5	2.0	Walk	12_01_23-DJI_0994	0 days 00:00:05

In [9]:

# zebras_sorted = zebras_sorted[zebras_sorted['behaviour'] != 'Out of Frame']

zebras = zebras[zebras['behaviour'] != 'Out of Frame']
zebras = zebras[zebras['behaviour']!= 'Occluded']
zebras = zebras[zebras['behaviour'] != 'Out of Focus']

# zebras_sorted = zebras_sorted[zebras_sorted['behaviour'] != 'Out of Frame'] zebras = zebras[zebras['behaviour'] != 'Out of Frame'] zebras = zebras[zebras['behaviour']!= 'Occluded'] zebras = zebras[zebras['behaviour'] != 'Out of Focus']

In [10]:

len(zebras)

len(zebras)

Out[10]:

933190

In [11]:

zebras['time'].max()

zebras['time'].max()

Out[11]:

Timedelta('0 days 02:25:22')

In [8]:

# combine behavior categories into stand-graze, walk, trot, and run
zebras['behaviour'] = zebras['behaviour'].replace('Graze', 'Graze')
zebras['behaviour'] = zebras['behaviour'].replace('Walk', 'Walk')
zebras['behaviour'] = zebras['behaviour'].replace('Trotting', 'Trot')
zebras['behaviour'] = zebras['behaviour'].replace('Running', 'Run')

# drop all other categories of behavior
zebras = zebras[zebras['behaviour'].isin(['Graze', 'Walk', 'Trot', 'Run'])]

# combine behavior categories into stand-graze, walk, trot, and run zebras['behaviour'] = zebras['behaviour'].replace('Graze', 'Graze') zebras['behaviour'] = zebras['behaviour'].replace('Walk', 'Walk') zebras['behaviour'] = zebras['behaviour'].replace('Trotting', 'Trot') zebras['behaviour'] = zebras['behaviour'].replace('Running', 'Run') # drop all other categories of behavior zebras = zebras[zebras['behaviour'].isin(['Graze', 'Walk', 'Trot', 'Run'])]

In [15]:

# downsample to 10 sec (30 fps)
zebras = zebras[zebras['time'].dt.seconds % (300) == 0]

# downsample to 10 sec (30 fps) zebras = zebras[zebras['time'].dt.seconds % (300) == 0]

In [16]:

# count each occurance of transitions between behaviours
# create dataframe to store the counts of transitions from one behaviour to another
behaviors = zebras['behaviour'].unique()
transition_counts = pd.DataFrame(index=behaviors, columns=behaviors)
transition_counts = transition_counts.fillna(0)

# count each occurance of transitions between behaviours # create dataframe to store the counts of transitions from one behaviour to another behaviors = zebras['behaviour'].unique() transition_counts = pd.DataFrame(index=behaviors, columns=behaviors) transition_counts = transition_counts.fillna(0)

In [17]:

# count the occurance of each transition for each unique zebra (id)
for video in zebras['video'].unique():
    for zebra in zebras[zebras['video'] == video]['id'].unique():
        zebra_data = zebras[(zebras['video'] == video) & (zebras['id'] == zebra)]
        for i in range(1, len(zebra_data)):
            transition_counts.loc[zebra_data.iloc[i-1]['behaviour'], zebra_data.iloc[i]['behaviour']] += 1

# count the occurance of each transition for each unique zebra (id) for video in zebras['video'].unique(): for zebra in zebras[zebras['video'] == video]['id'].unique(): zebra_data = zebras[(zebras['video'] == video) & (zebras['id'] == zebra)] for i in range(1, len(zebra_data)): transition_counts.loc[zebra_data.iloc[i-1]['behaviour'], zebra_data.iloc[i]['behaviour']] += 1

In [16]:

transition_counts

transition_counts

Out[16]:

	Walk	Graze	Run	Trot
Walk	650	97	16	20
Graze	95	1004	0	2
Run	8	0	26	4
Trot	15	3	4	6

In [17]:

# normalize transitions by the total number of transitions
transition_counts_normalized = transition_counts.div(transition_counts.sum(axis=1), axis=0)
transition_counts_normalized

# normalize transitions by the total number of transitions transition_counts_normalized = transition_counts.div(transition_counts.sum(axis=1), axis=0) transition_counts_normalized

Out[17]:

	Walk	Graze	Run	Trot
Walk	0.830140	0.123883	0.020434	0.025543
Graze	0.086285	0.911898	0.000000	0.001817
Run	0.210526	0.000000	0.684211	0.105263
Trot	0.535714	0.107143	0.142857	0.214286

In [20]:

# convert to percentages and rounds to 2 decimal places
transition_counts_normalized =round(transition_counts_normalized * 100, 2)

# convert to percentages and rounds to 2 decimal places transition_counts_normalized =round(transition_counts_normalized * 100, 2)

In [21]:

transition_counts_normalized
# switch order of walk and stand-graze in the table
transition_counts_normalized = transition_counts_normalized[['Graze', 'Walk', 'Trot', 'Run']]
#transition_counts_normalized = transition_counts_normalized.reindex(['Graze', 'Walk', 'Trot', 'Run'])

transition_counts_normalized # switch order of walk and stand-graze in the table transition_counts_normalized = transition_counts_normalized[['Graze', 'Walk', 'Trot', 'Run']] #transition_counts_normalized = transition_counts_normalized.reindex(['Graze', 'Walk', 'Trot', 'Run'])

In [22]:

# calculate the percentages row-wise for transitions
transition_percentages = transition_counts.div(transition_counts.sum(axis=1), axis=0)
# sort rows and columns by the order of the behaviors
behaviors = transition_percentages.columns
transition_percentages = transition_percentages.reindex(behaviors, axis=0)
transition_percentages = transition_percentages.reindex(behaviors, axis=1)
transition_percentages = round(transition_percentages*100, 2)

# calculate the percentages row-wise for transitions transition_percentages = transition_counts.div(transition_counts.sum(axis=1), axis=0) # sort rows and columns by the order of the behaviors behaviors = transition_percentages.columns transition_percentages = transition_percentages.reindex(behaviors, axis=0) transition_percentages = transition_percentages.reindex(behaviors, axis=1) transition_percentages = round(transition_percentages*100, 2)

In [24]:

# Reorder rows and columns to switch "Walk" and "Graze"
reordered_behaviors = ['Graze', 'Walk', 'Trot', 'Run']
transition_percentages = transition_percentages.reindex(reordered_behaviors, axis=0)
transition_percentages = transition_percentages.reindex(reordered_behaviors, axis=1)

# Plot the heatmap with the updated order
plt.figure(figsize=(10, 8))
sns.heatmap(transition_percentages, cmap='rocket', annot=True, fmt='g')

# Add numbers on each cell
for i in range(len(reordered_behaviors)):
    for j in range(len(reordered_behaviors)):
        plt.text(j + 0.5, i + 0.5, transition_percentages.iloc[i, j],
                 horizontalalignment='center',
                 verticalalignment='center')

plt.title('Transition Percentages Between Behaviors')
plt.xlabel('Current Behavior')
plt.ylabel('Previous Behavior')
plt.show()

# Reorder rows and columns to switch "Walk" and "Graze" reordered_behaviors = ['Graze', 'Walk', 'Trot', 'Run'] transition_percentages = transition_percentages.reindex(reordered_behaviors, axis=0) transition_percentages = transition_percentages.reindex(reordered_behaviors, axis=1) # Plot the heatmap with the updated order plt.figure(figsize=(10, 8)) sns.heatmap(transition_percentages, cmap='rocket', annot=True, fmt='g') # Add numbers on each cell for i in range(len(reordered_behaviors)): for j in range(len(reordered_behaviors)): plt.text(j + 0.5, i + 0.5, transition_percentages.iloc[i, j], horizontalalignment='center', verticalalignment='center') plt.title('Transition Percentages Between Behaviors') plt.xlabel('Current Behavior') plt.ylabel('Previous Behavior') plt.show()

In [27]:

# heatmap of transition percentages
plt.figure(figsize=(10, 8))
sns.heatmap(transition_percentages, cmap='cividis', fmt='g')
plt.title('Transitions Between Locomotion Behavior Categories (1 second intervals)')
plt.xticks(rotation=45, fontsize=15)
plt.yticks(rotation=0, fontsize=15)
plt.title('Transitions Between Zebra Locomotion Behaviors', fontsize=20)
plt.xlabel('Current Behavior', fontsize=20)
plt.ylabel('Previous Behavior', fontsize=20)
plt.show()

# heatmap of transition percentages plt.figure(figsize=(10, 8)) sns.heatmap(transition_percentages, cmap='cividis', fmt='g') plt.title('Transitions Between Locomotion Behavior Categories (1 second intervals)') plt.xticks(rotation=45, fontsize=15) plt.yticks(rotation=0, fontsize=15) plt.title('Transitions Between Zebra Locomotion Behaviors', fontsize=20) plt.xlabel('Current Behavior', fontsize=20) plt.ylabel('Previous Behavior', fontsize=20) plt.show()

In [ ]: