elo-rank-photos/rank_photos.py

427 lines
9.3 KiB
Python
Raw Normal View History

2016-02-20 20:52:43 +01:00
#!/usr/bin/env python
"""
Matplotlib based photo ranking system using the Elo rating system.
Reference: http://en.wikipedia.org/wiki/Elo_rating_system
by Nick Hilton
This file is in the public domain.
"""
# Python
import argparse
import glob
import json
import os
import sys
# 3rd party
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.image as mpimg
import exifread
class Photo:
LEFT = 0
RIGHT = 1
def __init__(self, filename, score = 1400.0, wins = 0, matches = 0):
if not os.path.isfile(filename):
raise ValueError("Could not find the file: %s" % filename)
self._filename = filename
self._score = score
self._wins = wins
self._matches = matches
# read orientation with exifread
with open(filename, 'rb') as fd:
tags = exifread.process_file(fd)
self._rotation = str(tags['Image Orientation'])
def filename(self):
return self._filename
def matches(self):
return self._matches
def rotation(self):
return self._rotation
def score(self, s = None, is_winner = None):
if s is None:
return self._score
assert is_winner is not None
self._score = s
self._matches += 1
if is_winner:
self._wins += 1
def win_percentage(self):
return 100.0 * float(self._wins) / float(self._matches)
def __eq__(self, rhs):
return self._filename == rhs._filename
def to_dict(self):
return {
'filename' : self._filename,
'score' : self._score,
'matches' : self._matches,
'wins' : self._wins,
}
class Display(object):
"""
Given two photos, displays them with Matplotlib and provides a graphical
means of choosing the better photo.
"""
def __init__(self, f1, f2, title = None, figsize = None):
self._choice = None
assert isinstance(f1, Photo)
assert isinstance(f2, Photo)
if figsize is None:
figsize = [20,12]
fig = plt.figure(figsize=figsize)
h = 10
ax11 = plt.subplot2grid((h,2), (0,0), rowspan = h - 1)
ax12 = plt.subplot2grid((h,2), (0,1), rowspan = h - 1)
ax21 = plt.subplot2grid((h,6), (h - 1, 1))
ax22 = plt.subplot2grid((h,6), (h - 1, 4))
kwargs = dict(s = 'Select', ha = 'center', va = 'center', fontsize=20)
ax21.text(0.5, 0.5, **kwargs)
ax22.text(0.5, 0.5, **kwargs)
self._fig = fig
self._ax_select_left = ax21
self._ax_select_right = ax22
fig.subplots_adjust(
left = 0.02,
bottom = 0.02,
right = 0.98,
top = 0.98,
wspace = 0.05,
hspace = 0,
)
left = self._read(f1)
right = self._read(f2)
ax11.imshow(left)
ax12.imshow(right)
for ax in [ax11, ax12, ax21, ax22]:
ax.set_xticklabels([])
ax.set_yticklabels([])
ax.set_xticks([])
ax.set_yticks([])
self._attach_callbacks()
if title:
fig.suptitle(title, fontsize=20)
plt.show()
def _read(self, photo):
data = mpimg.imread(photo.filename())
r = photo.rotation()
#~ print "Rotation: ", r
#~ print " data.shape = ", data.shape
# rotate as necessary
if r == 'Horizontal (normal)':
pass
elif r == 'Rotated 90 CW':
data = np.rot90(data, 3)
elif r == 'Rotated 90 CCW':
data = np.rot90(data, 1)
else:
raise RuntimeError('Unhandled rotation "%s"' % r)
#~ print " data.shape = ", data.shape
return data
def _on_click(self, event):
if event.inaxes == self._ax_select_left:
self._choice = Photo.LEFT
plt.close(self._fig)
elif event.inaxes == self._ax_select_right:
self._choice = Photo.RIGHT
plt.close(self._fig)
def _attach_callbacks(self):
self._fig.canvas.mpl_connect('button_press_event', self._on_click)
class EloTable:
def __init__(self, max_increase = 32.0):
self._K = max_increase
self._photos = {}
self._shuffled_keys = []
def add_photo(self, photo):
filename = photo.filename()
if filename not in self._photos:
self._photos[filename] = photo
def get_ranked_list(self):
# Convert the dictionary into a list and then sort by score.
ranked_list = self._photos.values()
ranked_list = sorted(
ranked_list,
key = lambda record : record.score(),
reverse = True)
return ranked_list
def rank_photos(self, n_iterations, figsize):
"""
Displays two photos using the command "gnome-open". Then asks which
photo is better.
"""
n_photos = len(self._photos)
keys = self._photos.keys()
for i in xrange(n_iterations):
np.random.shuffle(keys)
n_matchups = n_photos / 2
for j in xrange(0, n_photos, 2):
match_up = j / 2
title = 'Round %d / %d, Match Up %d / %d' % (
i + 1, n_iterations,
match_up + 1,
n_matchups)
photo_a = self._photos[keys[j]]
photo_b = self._photos[keys[j+1]]
d = Display(photo_a, photo_b, title, figsize)
if d == Photo.LEFT:
self.__score_result(photo_a, photo_b)
else:
self.__score_result(photo_b, photo_a)
def __score_result(self, winning_photo, loosing_photo):
# Current ratings
R_a = winning_photo.score()
R_b = loosing_photo.score()
# Expectation
E_a = 1.0 / (1.0 + 10.0 ** ((R_a - R_b) / 400.0))
E_b = 1.0 / (1.0 + 10.0 ** ((R_b - R_a) / 400.0))
# New ratings
R_a = R_a + self._K * (1.0 - E_a)
R_b = R_b + self._K * (0.0 - E_b)
winning_photo.score(R_a, True)
loosing_photo.score(R_b, False)
def to_dict(self):
rl = self.get_ranked_list()
rl = [x.to_dict() for x in rl]
return {'photos' : rl}
def main():
description = """\
Uses the Elo ranking algorithm to sort your images by rank. The program reads
the comand line for images to present to you in random order, then you select
the better photo. After N iteration the resulting rankings are displayed.
"""
parser = argparse.ArgumentParser(description = description)
parser.add_argument(
"-r",
"--n-rounds",
type = int,
default = 3,
help = "Specifies the number of rounds to pass through the photo set"
)
parser.add_argument(
"-f",
"--figsize",
nargs = 2,
type = int,
default = [20, 12],
help = "Specifies width and height of the Matplotlib figsize (20, 12)"
)
parser.add_argument(
"photo_dir",
help = "The photo directory to scan for .jpg images"
)
args = parser.parse_args()
assert os.path.isdir(args.photo_dir)
os.chdir(args.photo_dir)
ranking_table_json = 'ranking_table.json'
ranked_txt = 'ranked.txt'
# Create the ranking table and add photos to it.
table = EloTable()
#--------------------------------------------------------------------------
# Read in table .json if present
if os.path.isfile(ranking_table_json):
with open(ranking_table_json, 'r') as fd:
d = json.load(fd)
# create photos and add to table
for p in d['photos']:
photo = Photo(**p)
table.add_photo(photo)
#--------------------------------------------------------------------------
# glob for files, to include newly added files
filelist = glob.glob('*.jpg')
photos = [Photo(x) for x in filelist]
for p in photos:
table.add_photo(p)
#--------------------------------------------------------------------------
# Rank the photos!
table.rank_photos(args.n_rounds, args.figsize)
#--------------------------------------------------------------------------
# save the table
with open(ranking_table_json, 'w') as fd:
d = table.to_dict()
jstr = json.dumps(d, indent = 4, separators=(',', ' : '))
fd.write(jstr)
#--------------------------------------------------------------------------
# dump ranked list to disk
with open(ranked_txt, 'w') as fd:
ranked_list = table.get_ranked_list()
heading_fmt = "%4d %4.0f %7d %7.2f %s\n"
heading = "Rank Score Matches Win % Filename\n"
fd.write(heading)
for i, photo in enumerate(ranked_list):
line = heading_fmt %(
i + 1,
photo.score(),
photo.matches(),
photo.win_percentage(),
photo.filename())
fd.write(line)
#--------------------------------------------------------------------------
# dump ranked list to screen
print "Final Ranking:"
with open(ranked_txt, 'r') as fd:
text = fd.read()
print text
if __name__ == "__main__": main()