-
Notifications
You must be signed in to change notification settings - Fork 17
/
calibrateObjective.py
130 lines (114 loc) · 4.61 KB
/
calibrateObjective.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
from learnedRanking import learnToRank
from distanceMetrics import *
from fastRender import fastRender
from recognitionModel import Particle,RecognitionModel
from groundTruthParses import groundTruth
from utilities import *
import tensorflow as tf
import numpy as np
import os
import pickle
MODE = 'ranking' # distance
class DummyArguments():
def __init__(self):
self.noisy = True
self.distance = True
self.architecture = "original"
self.dropout = False
self.learningRate = 1
self.showParticles = False
worker = RecognitionModel(DummyArguments())
worker.loadDistanceCheckpoint("checkpoints/distance.checkpoint")
def featuresOfParticle(p, target):
f = map(float,[p.logLikelihood, p.program.logPrior(),
p.distance[0],p.distance[1]])
return f
# -0.01*asymmetricBlurredDistance(target,fastRender(p.program),
# kernelSize = 7,
# factor = 1,
# invariance = 3)]
distanceTrainingData = []
trainingData = []
print groundTruth.keys()
for k in groundTruth:
target = loadImage(k)
print k
parseDirectory = k[:k.index('.')] + '-parses/'
negatives = []
positives = []
for p in range(300):
pkl = parseDirectory + 'particle%d.p'%p
if not os.path.isfile(pkl):
break
particle = pickle.load(open(pkl,'rb'))
print " [+] Loaded %s"%pkl
particle.output = None
if set(map(str,particle.program.lines)) == groundTruth[k]:
positives.append(particle)
else:
negatives.append(particle)
print "Got %d positive examples"%(len(positives))
if len(positives) > 0 and len(negatives) > 0:
if MODE == 'ranking':
for p in positives + negatives: p.output = fastRender(p.program)
worker.learnedParticleDistances(target,positives + negatives)
trainingData.append((np.array(map(lambda p: featuresOfParticle(p,target),positives)),
np.array(map(lambda p: featuresOfParticle(p,target),negatives))))
else:
distanceTrainingData.append((target,
[ fastRender(p.program) for p in positives ],
[ fastRender(p.program) for p in negatives ]))
if MODE == 'distance':
print "Calibrating distance function..."
def ranks(kernelSize, factor, invariance):
rs = []
for t,positives,negatives in distanceTrainingData:
positiveScores = [ asymmetricBlurredDistance(t,p,kernelSize = kernelSize,factor = factor,invariance = invariance)
for p in positives ]
negativeScores = [ asymmetricBlurredDistance(t,p,kernelSize = kernelSize,factor = factor,invariance = invariance)
for p in negatives ]
bestPositive = min(positiveScores)
rs.append(len([ None for n in negativeScores if n <= bestPositive ]) + 1)
return rs
for k in [1,3,5,7]:
for i in [0,1,2,3]:
for f in [0.5,1,2,5,10]:
rs = ranks(k,f,i)
print (k,i,f),
print len([ None for r in rs if r < 2 ]),
print len([ None for r in rs if r < 5+1 ]),
print len([ None for r in rs if r < 10+1 ])
assert False
parameters = learnToRank(trainingData)
def ranks(w):
rs = []
for positives, negatives in trainingData:
bestPositive = np.dot(positives,w).max()
negativeScores = np.dot(negatives,w)
rs.append((negativeScores > bestPositive).sum() + 1)
return rs
# evaluate learned parameters
rs = ranks(parameters)
print "Top 1/5/10:"
print len([r for r in rs if r == 1 ])
print len([r for r in rs if r < 6 ])
print len([r for r in rs if r < 11 ])
print
topTen = {}
for priorWeight in range(0,30):
for distanceWeight in range(0,30):
w = np.array([1,priorWeight/20.0,distanceWeight/200.0])
rs = ranks(w)
print w
print len([r for r in rs if r < 11 ])
topTen[tuple(w.tolist())] = len([r for r in rs if r < 11 ])
print "\n".join(map(str,list(sorted(topTen.items(),key = lambda kv: kv[1]))))
# use the weights they give us the best in the top ten
w,_ = max(topTen.items(),key = lambda kv: kv[1])
w = np.array(w)
r = ranks(w)
print r
for j in range(1,50):
print "Top",j,":", len([x for x in r if x < j+1 ])
print "# examples:",len(r)
print "Average rank:",sum(r)/float(len(r))