annotate analysis_of_annotations/evaluate_annotations.py @ 5:b523456082ca tip

Update path to dataset and reflect modified chunk naming convention.

#!/usr/bin/python
#
# evaluate_annotations.py:
#    Compute descriptive statistics of annotations, including annotator
#    agreement 
#
# Author: Peter Foster
# (c) 2015 Peter Foster
#

from pandas import Series, DataFrame
import glob
import pandas.io.parsers
from pylab import *
import numpy as np
import re

DatasetPath = '/import/c4dm-02/people/peterf/audex/datasets/chime_home/'

#Read in annotations
Chunks = list(Series.from_csv(DatasetPath + 'release_chunks_raw.csv',header=None))
Annotations = []
for chunk in Chunks: 
    Annotations.append(Series.from_csv(DatasetPath + 'chunks/' + chunk + '.csv'))
Annotations = DataFrame(Annotations)

Annotators = ('annotation_a1', 'annotation_a2', 'annotation_a3')

#Check label integrity
Labels = set('cmfvpboSU')
for a in Annotators:
    I = Annotations[a].notnull()
    Annotations[a].ix[I] = Annotations[a][I].apply(lambda s: ''.join(sorted(s)))
    assert(all(Annotations[a].ix[I].apply(lambda s: len(set(s).difference(Labels)) == 0)))    

#Annotator-wise label counts
CountStats = {'annotatorwise':{}, 'majorityvote':{}}
for annotator in Annotators:
    CountStats['annotatorwise'][annotator] = {}
    for label in Labels:
        V1 = Annotations[annotator][Annotations[annotator].notnull()]
        V1 = V1.apply(lambda s: label in s)
        CountStats['annotatorwise'][annotator][label] = sum(V1)
CountStats['annotatorwise'] = DataFrame(CountStats['annotatorwise'])
#Rearrange index for plotting histogram
CountStats['annotatorwise'] = CountStats['annotatorwise'].ix[['c','m','f','v','p','b','o','S','U']]
CountStats['annotatorwise_coefficient_of_variation'] = (CountStats['annotatorwise'].T.std() / CountStats['annotatorwise'].T.mean())
CountStats['annotatorwise_coefficient_of_variation'].sort()

#Histogram of label counts
fig_width_pt = 246.0 # Get this from LaTeX using \showthe\columnwidth
inches_per_pt = 1.0/72.27 # Convert pt to inch
golden_mean = (sqrt(5)-1.0)/2.0 # Aesthetic ratio
fig_width = fig_width_pt*inches_per_pt # width in inches
fig_height = fig_width*golden_mean # height in inches
fig_size =  [fig_width,fig_height]
params = {'backend': 'ps',
            'axes.labelsize': 8,
            'text.fontsize': 8,
            'legend.fontsize': 7.0,
            'xtick.labelsize': 8,
            'ytick.labelsize': 8,
            'text.usetex': False,
            'figure.figsize': fig_size}
rcParams.update(params)
ind = np.arange(len(CountStats['annotatorwise']))  # the x locations for the groups
width = 0.29       # the width of the bars
fig, ax = plt.subplots()
rects = []
colours = ('r', 'y', 'g')
for annotator, i in zip(Annotators, range(len(Annotators))):
    rects.append(ax.bar(ind+width*i, CountStats['annotatorwise'][annotator], width, color=colours[i], align='center'))
# add text for labels, title and axes ticks
ax.set_ylabel('Count')
ax.set_xlabel('Label')
#ax.set_title('Annotator-wise label histogram')
ax.set_xticks(ind+width)
ax.set_xticklabels(CountStats['annotatorwise'].index)
ax.legend( (rect[0] for rect in rects), ('Annotator 1', 'Annotator 2', 'Annotator 3') )
#Tweak x-axis limit
ax.set_xlim(left=-0.5)
ax.set_ylim(top=3500)
plt.gcf().subplots_adjust(left=0.15) #Prevent y-axis label from being chopped off
def autolabel(r):
    for rects in r:
        for rect in rects:
            height = rect.get_height()
            ax.text(rect.get_x()+0.180,100+height,'%d'%int(height),ha='center',va='bottom',rotation='vertical',size=6.0)
autolabel(rects)
plt.draw()
plt.savefig('figures/annotation_histogram.pdf')

#Generalised Jaccard index
def jaccardIndex(*Sets):
    SetN = Sets[0]
    SetD = Sets[0]
    for S in Sets:
        SetN = SetN.intersection(S)
        SetD = SetD.union(S)
    if len(SetD) == 0:
        return 1
    else:
        return len(SetN) / float(len(SetD))        

AgreementStats = {'jaccardindex_pos':{}}
for l in Labels:
    V1 = set(np.where(Annotations['annotation_a1'].apply(lambda s: l in s))[0])
    V2 = set(np.where(Annotations['annotation_a2'].apply(lambda s: l in s))[0])
    V3 = set(np.where(Annotations['annotation_a3'].apply(lambda s: l in s))[0])   
    AgreementStats['jaccardindex_pos'][l] = {}
    AgreementStats['jaccardindex_pos'][l]['a1_a2_a3'] = jaccardIndex(V1, V2, V3)
    AgreementStats['jaccardindex_pos'][l]['a1_a2'] = jaccardIndex(V1, V2)
    AgreementStats['jaccardindex_pos'][l]['a2_a3'] = jaccardIndex(V2, V3)
    AgreementStats['jaccardindex_pos'][l]['a1_a3'] = jaccardIndex(V1, V3)
#Experiment with combining label classes
V1 = set(np.where(Annotations['annotation_a1'].apply(lambda s: 'o' in s or 'p' in s))[0])
V2 = set(np.where(Annotations['annotation_a2'].apply(lambda s: 'o' in s or 'p' in s))[0])
V3 = set(np.where(Annotations['annotation_a3'].apply(lambda s: 'o' in s or 'p' in s))[0])
l = '{o,p}'
AgreementStats['jaccardindex_pos'][l] = {}
AgreementStats['jaccardindex_pos'][l]['a1_a2_a3'] = jaccardIndex(V1, V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a2'] = jaccardIndex(V1, V2)
AgreementStats['jaccardindex_pos'][l]['a2_a3'] = jaccardIndex(V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a3'] = jaccardIndex(V1, V3)
l = '{o,b}'
V1 = set(np.where(Annotations['annotation_a1'].apply(lambda s: 'o' in s or 'b' in s))[0])
V2 = set(np.where(Annotations['annotation_a2'].apply(lambda s: 'o' in s or 'b' in s))[0])
V3 = set(np.where(Annotations['annotation_a3'].apply(lambda s: 'o' in s or 'b' in s))[0]) 
AgreementStats['jaccardindex_pos'][l] = {}
AgreementStats['jaccardindex_pos'][l]['a1_a2_a3'] = jaccardIndex(V1, V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a2'] = jaccardIndex(V1, V2)
AgreementStats['jaccardindex_pos'][l]['a2_a3'] = jaccardIndex(V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a3'] = jaccardIndex(V1, V3)
l = '{o,S}'
V1 = set(np.where(Annotations['annotation_a1'].apply(lambda s: 'o' in s or 'S' in s))[0])
V2 = set(np.where(Annotations['annotation_a2'].apply(lambda s: 'o' in s or 'S' in s))[0])
V3 = set(np.where(Annotations['annotation_a3'].apply(lambda s: 'o' in s or 'S' in s))[0]) 
AgreementStats['jaccardindex_pos'][l] = {}
AgreementStats['jaccardindex_pos'][l]['a1_a2_a3'] = jaccardIndex(V1, V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a2'] = jaccardIndex(V1, V2)
AgreementStats['jaccardindex_pos'][l]['a2_a3'] = jaccardIndex(V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a3'] = jaccardIndex(V1, V3)

AgreementStats['jaccardindex_pos'] = DataFrame(AgreementStats['jaccardindex_pos']).T
AgreementStats['jaccardindex_pos'] = AgreementStats['jaccardindex_pos'][['a1_a2', 'a1_a3', 'a2_a3', 'a1_a2_a3']]
print('Agreement about label presence (unfiltered dataset)')
print(AgreementStats['jaccardindex_pos'].ix[['c','m','f','v','p','b','o','S','{o,p}','{o,b}','{o,S}']])
#Coefficients of variation across pairs of annotators
A = (AgreementStats['jaccardindex_pos'].ix[['c','m','f','v','p','b','o','S','{o,p}','{o,b}','{o,S}']][['a1_a2', 'a1_a3', 'a2_a3']].T)
print('Coefficients of variation')
print(A.std() / A.mean())

#Read in annotations for refined dataset
Chunks = list(Series.from_csv(DatasetPath + 'release_chunks_refined.csv',header=None))
Annotations = []
for chunk in Chunks: 
    Annotations.append(Series.from_csv(DatasetPath + 'chunks/' + chunk + '.csv'))
Annotations = DataFrame(Annotations)

AgreementStats = {'jaccardindex_pos':{}}
for l in Labels:
    V1 = set(np.where(Annotations['annotation_a1'].apply(lambda s: l in s))[0])
    V2 = set(np.where(Annotations['annotation_a2'].apply(lambda s: l in s))[0])
    V3 = set(np.where(Annotations['annotation_a3'].apply(lambda s: l in s))[0]) 

    AgreementStats['jaccardindex_pos'][l] = {}
    AgreementStats['jaccardindex_pos'][l]['a1_a2_a3'] = jaccardIndex(V1, V2, V3)
    AgreementStats['jaccardindex_pos'][l]['a1_a2'] = jaccardIndex(V1, V2)
    AgreementStats['jaccardindex_pos'][l]['a2_a3'] = jaccardIndex(V2, V3)
    AgreementStats['jaccardindex_pos'][l]['a1_a3'] = jaccardIndex(V1, V3)
#Experiment with combining label classes
V1 = set(np.where(Annotations['annotation_a1'].apply(lambda s: 'o' in s or 'p' in s))[0])
V2 = set(np.where(Annotations['annotation_a2'].apply(lambda s: 'o' in s or 'p' in s))[0])
V3 = set(np.where(Annotations['annotation_a3'].apply(lambda s: 'o' in s or 'p' in s))[0])
l = '{o,p}'
AgreementStats['jaccardindex_pos'][l] = {}
AgreementStats['jaccardindex_pos'][l]['a1_a2_a3'] = jaccardIndex(V1, V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a2'] = jaccardIndex(V1, V2)
AgreementStats['jaccardindex_pos'][l]['a2_a3'] = jaccardIndex(V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a3'] = jaccardIndex(V1, V3)
l = '{o,b}'
V1 = set(np.where(Annotations['annotation_a1'].apply(lambda s: 'o' in s or 'b' in s))[0])
V2 = set(np.where(Annotations['annotation_a2'].apply(lambda s: 'o' in s or 'b' in s))[0])
V3 = set(np.where(Annotations['annotation_a3'].apply(lambda s: 'o' in s or 'b' in s))[0]) 
AgreementStats['jaccardindex_pos'][l] = {}
AgreementStats['jaccardindex_pos'][l]['a1_a2_a3'] = jaccardIndex(V1, V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a2'] = jaccardIndex(V1, V2)
AgreementStats['jaccardindex_pos'][l]['a2_a3'] = jaccardIndex(V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a3'] = jaccardIndex(V1, V3)
l = '{o,S}'
V1 = set(np.where(Annotations['annotation_a1'].apply(lambda s: 'o' in s or 'S' in s))[0])
V2 = set(np.where(Annotations['annotation_a2'].apply(lambda s: 'o' in s or 'S' in s))[0])
V3 = set(np.where(Annotations['annotation_a3'].apply(lambda s: 'o' in s or 'S' in s))[0]) 
AgreementStats['jaccardindex_pos'][l] = {}
AgreementStats['jaccardindex_pos'][l]['a1_a2_a3'] = jaccardIndex(V1, V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a2'] = jaccardIndex(V1, V2)
AgreementStats['jaccardindex_pos'][l]['a2_a3'] = jaccardIndex(V2, V3)
AgreementStats['jaccardindex_pos'][l]['a1_a3'] = jaccardIndex(V1, V3)

AgreementStats['jaccardindex_pos'] = DataFrame(AgreementStats['jaccardindex_pos']).T
AgreementStats['jaccardindex_pos'] = AgreementStats['jaccardindex_pos'][['a1_a2', 'a1_a3', 'a2_a3', 'a1_a2_a3']]
print('Agreement about label presence (refined dataset)')
print(AgreementStats['jaccardindex_pos'].ix[['c','m','f','v','p','b','o','S','{o,p}','{o,b}','{o,S}']])
#Coefficients of variation across pairs of annotators
A = (AgreementStats['jaccardindex_pos'].ix[['c','m','f','v','p','b','o','S','{o,p}','{o,b}','{o,S}']][['a1_a2', 'a1_a3', 'a2_a3']].T)
print('Coefficients of variation')
print(A.std() / A.mean())

#Label count stats for majority vote
CountStats['majorityvote'] = {}
for label in 'cmfvpboS':
    CountStats['majorityvote'][label] = {}
    for comparisonLabel in 'cmfvpboS':
        V1 = Annotations['majorityvote']
        V1 = V1.apply(lambda s: label in s and comparisonLabel in s)
        CountStats['majorityvote'][label][comparisonLabel] = sum(V1)
CountStats['majorityvote'] = DataFrame(CountStats['majorityvote'])
print('Label co-occurrences')
CountStats['majorityvote'].loc[['c','m','f','v','p','b','o','S'],['c','m','f','v','p','b','o','S']]