Skip to content

Commit

Permalink
external output
Browse files Browse the repository at this point in the history
  • Loading branch information
@abhimanyudubey
abhimanyudubey committed Aug 24, 2016
1 parent 3a210bf commit b26f75d
Show file tree
Hide file tree
Showing 2 changed files with 236 additions and 1 deletion.
235 changes: 235 additions & 0 deletions code/vis_generate_comparisons.py
View file
Original file line number Diff line number Diff line change
@@ -0,0 +1,235 @@
import caffe
import sys
import glob
import random
import os
from trueskill import TrueSkill
import numpy as np
from multiprocessing import Pool


def load_image(inp):
inp_file = inp[0]
transformer = inp[1]
image = transformer.preprocess('data', caffe.io.load_image(inp_file))
return image


def generate_scores(voteFile):
# function to generate scores using TrueSkill using a vote file given in the format:
# image1 image2 winning-image

f1 = open(voteFile);

mu0 = 25;
var0 = (mu0 / 3.0) ** 2;
beta2 = var0 / 4.0;
tau2 = var0 / 100.0;
prob_draw = 0.1333;
ts = TrueSkill(beta2, tau2, prob_draw, var0);
params = {};
vals = [[], [], [], [], []];
c = 0;

for vote in f1:
vote = vote.strip().split();
win = 0;
if vote[1] == vote[2]: win = 1;
if vote[0] not in params: params[vote[0]] = (mu0, var0);
if vote[1] not in params: params[vote[1]] = (mu0, var0);
if not win:
params[vote[0]], params[vote[1]] = ts.update_rating(params[vote[0]], params[vote[1]], False);
else:
params[vote[1]], params[vote[0]] = ts.update_rating(params[vote[1]], params[vote[0]], False);
if c % 10000 == 0:
mu_mu, mu_sigma, sigma_mu, sigma_sigma = 0.0, 0.0, 0.0, 0.0;
mus, sigmas = [], [];

for key, value in params.iteritems():
mu_mu += value[0];
sigma_mu += value[1];
mus.append(value[0]);
sigmas.append(value[1]);

mu_mu = np.mean(mus);
sigma_mu = np.mean(sigmas);
mu_sigma = np.std(np.array(mus));
sigma_sigma = np.std(np.array(sigmas));

vals[0].append(mu_mu);
vals[1].append(sigma_mu);
vals[2].append(mu_sigma);
vals[3].append(sigma_sigma);
vals[4].append(c);

c += 1;

f1.close();

return params, ts, vals;


def update_scores(voteFile, ts, params):
# function to update scores given an existing trueskill object and parameters using a different votefile
mu0 = 25;
var0 = (mu0 / 3.0) ** 2;
beta2 = var0 / 4.0;
tau2 = var0 / 100.0;
prob_draw = 0.1333;
f1 = open(voteFile);
c = 0;
vals = [[], [], [], [], []];

for vote in f1, f2:
vote = vote.replace(',', ' ');
vote = vote.strip().split();
win = 0;
if vote[1] == vote[2]: win = 1;
if vote[0] not in params: params[vote[0]] = (mu0, var0);
if vote[1] not in params: params[vote[1]] = (mu0, var0);
if not win:
params[vote[0]], params[vote[1]] = ts.update_rating(params[vote[0]], params[vote[1]], False);
else:
params[vote[1]], params[vote[0]] = ts.update_rating(params[vote[1]], params[vote[0]], False);
if c % 10000 == 0:
mu_mu, mu_sigma, sigma_mu, sigma_sigma = 0.0, 0.0, 0.0, 0.0;
mus, sigmas = [], [];

for key, value in params.iteritems():
mu_mu += value[0];
sigma_mu += value[1];
mus.append(value[0]);
sigmas.append(value[1]);

mu_mu = np.mean(mus);
sigma_mu = np.mean(sigmas);
mu_sigma = np.std(np.array(mus));
sigma_sigma = np.std(np.array(sigmas));

vals[0].append(mu_mu);
vals[1].append(sigma_mu);
vals[2].append(mu_sigma);
vals[3].append(sigma_sigma);
vals[4].append(c);

c += 1;

f1.close();

return params, ts, vals;


if __name__ == "__main__":

if (len(sys.argv) < 5):
print "Usage python vis_generate_comparisons.py <folder-name> <output-folder-name>"
else:
dir_src = sys.argv[1]
dir_out = sys.argv[2]
# this script takes in a source directory and output directory and a
# caffe model and generate votes for all images present in all subfolders of this directory
file_ref = sys.argv[3]
dir_ref_images = sys.argv[4]

# first generating global reference parameters
global_params, global_ts, global_vals = generate_scores(file_ref)
global_image_dir = glob.glob(os.path.join(dir_ref_images, '*.jpg')) + glob.glob(
os.path.join(dir_ref_images, '*.JPG'))

subdir_src = [x for x in glob.glob(os.path.join(dir_src, '*')) if os.path.isdir(x)]

if not os.path.exists(os.path.join(dir_out, 'votes')):
os.makedirs(os.path.join(dir_out, 'votes'))

if not os.path.exists(os.path.join(dir_out, 'scores')):
os.makedirs(os.path.join(dir_out, 'scores'))

out_pairs = []

for subdir in subdir_src:
# for each subdirectory of images, we do the comparisons
# with 15 random images from reference along 15 random from set with their votes
list_img = glob.glob(os.path.join(subdir, '*.jpg')) + glob.glob(os.path.join(subdir, '*.JPG'))
file_vote = os.path.join(dir_out, 'votes', os.path.basename(subdir))
file_scores = os.path.join(dir_out, 'scores', os.path.basename(subdir))

out_images1 = [];
out_images2 = [];

print 'Generating comparisons for subdirectory', subdir, 'right now'
print 'Total number of images', len(list_img), ', total votes generated will be', len(list_img) * 20

for i in range(1, len(list_img) + 1):
# first 15 votes will be with random images and the image itself
img1 = list_img[i]

j_list1 = [];
for j in range(15):
j_list1.append(random.choice(global_image_dir))
for j in range(15):
j_list1.append(random.choice(list_img_copy))


for j in range(30):
out_pairs.append((img1, j_list1[j]))
# out_images1.append(image1)
# out_images2.append(res1[j])
#
#
# if i%batchsize==0 and i>0:
# pred = []
# out = net.forward_all(data1=np.asarray(out_images1),data2=np.asarray(out_images2))
# preds = out['fc8r']
#
# for k in range(0,batchsize*20):
# pred = preds[k] / np.linalg.norm(preds[k], ord=1)
# if pred[0] > pred[1]:
# votes.write(out_pairs[k][0] + ' ' + out_pairs[k][1] + ' ' + out_pairs[k][0] + '\n')
# else:
# votes.write(out_pairs[k][0] + ' ' + out_pairs[k][1] + ' ' + out_pairs[k][1] + '\n')
#
# out_images1 = []
# out_images2 = []
# out_pairs = []
#
# print 'Generated',i*20,'votes out of ',len(list_img)*20
#
# if len(out_pairs)>0:
# # tackling the left-over images
# pred = []
#
# while not len(out_images1)==batchsize*20:
# out_images1.append(out_images1[0])
# out_images2.append(out_images1[0])
#
# out = net.forward_all(data1=np.asarray(out_images1), data2=np.asarray(out_images2))
# preds = out['fc8r']
#
# for k in range(0, out_pairs):
# pred = preds[k] / np.linalg.norm(preds[k], ord=1)
# if pred[0] > pred[1]:
# votes.write(out_pairs[k][0] + ' ' + out_pairs[k][1] + ' ' + out_pairs[k][0] + '\n')
# else:
# votes.write(out_pairs[k][0] + ' ' + out_pairs[k][1] + ' ' + out_pairs[k][1] + '\n')
#
# out_images1 = []
# out_images2 = []
# out_pairs = []
#
# print 'Generated all votes in this subdirectory'
#
# votes.close()
# # now that votes are generated we will write them to a file after generating scores
# subdir_params,subdir_ts,subdir_vals = update_scores(file_vote,global_params,global_ts)
# for key,value in subdir_params.iteritems():
# if key in list_img:
# # write only those scores that belong to that subdir
# file_scores.write(','.join([key,str(value[0])]))
#
# file_scores.close()

f1 = open('/home/dubeya/extrapolated_votes/comparisons_1', 'w')
f2 = open('/home/dubeya/extrapolated_votes/comparisons_2', 'w')



2 changes: 1 addition & 1 deletion code/vis_generate_scores.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -170,7 +170,7 @@ def update_scores(voteFile, ts, params):
out_pairs = [];
out_images1 = [];
out_images2 = [];
batchsize = 60; # for 30 images this gives us 1110 (close to GPU capacity)
batchsize = 56; # for 30 images this gives us 1110 (close to GPU capacity)

print 'Generating votes for subdirectory',subdir, 'right now'
print 'Total number of images',len(list_img),', total votes generated will be',len(list_img)*20
Expand Down

0 comments on commit b26f75d

Please sign in to comment.