def biasMeasGenerator(generator: MeasurementGenerator, priors: Tuple[float],
conds: List[List[float]]):
# Updates measures with bias towards those indicated by class conditional probabilites.
# No output. Destructively alters generator.
newProbs = calcMeasProb(priors, conds)
generator.updateProbs(newProbs)
return
def main():
from random import seed
from probability import genProbs
#### Unit test for training funciton
# Initiatializes seed for recurrent testing.
for _ in range(10):
dimen = (4, 3, 6, 5, 6)
classValues = 5
measures = MeasurementGenerator(dimen)
classes = ClassAssign(dimen, classValues)
conds = [list(genProbs(measures.range)) for _ in range(classValues)]
egain = [[2, 0, 0, 0, 1], [3, 4, 0, 2, 2], [2, 2, 5, 1, 1],
[2, 2, 3, 4, 1], [0, 1, -3, 2, 3]]
classifier = BayesClassifier(
None, conds, eGain=egain
) # Worries that supplying similar priors is affecting our results. Even though vFold updates.
y = train(classifier, 20, measures, classes, 6000, delta=.0005)
z = [y[i] - y[i - 1] for i in range(1, len(y))]
# Trying to figure out average negative error to see if this is floating point.
print(y)
print()
print(z)
q = [i for i in z if i < 0]
q = sum(q) / max(len(q), 1)
print(q)
print()
x = measures.genMeas(20)
p = classes.assign(x)
l = classifier.assign(x)
def train(tagger: BayesClassifier,
iter: int,
measGen: MeasurementGenerator,
classGen: ClassAssign,
Z: int,
V: int = 10,
delta: float = .05) -> List[float]:
# Performs 'iter' iterations of vFold testing (default 'V' is ten) with 'tagger' classifier
# for 'Z' samples generated by 'measGen' and 'classGen.' After each vFold validation, appends
# an expected value (attached to tagger) and then optimizes tagger by 'delta' paramenter (default .05).
# Outputs a new optimized tagger and list of gain values from each iteration.
expectedGain = []
for _ in range(iter):
# Generates measurements
samples = measGen.genMeas(Z)
values = classGen.assign(samples)
# Shuffles values
samplesSh, valuesSh = shuffle(samples, values)
# Performs Test
matrix = vFold(samplesSh, valuesSh, V, tagger)
# Appends value to list
expectedGain.append(calcExpGain(matrix, tagger.eGain))
# Gives class probability over whole data set.
tagger.priorUpdate(calcClassProb(valuesSh, tagger.range))
# Updates tagger
tagger.optimize(delta, samplesSh, valuesSh)
return expectedGain
def main():
from bayes import calcClassProb
priors = (.6, .4)
conds = [[.12 / .6, .18 / .6, .3 / .6], [.2 / .4, .16 / .4, .04 / .4]]
gain = ((1, 0), (0, 2))
### Tests for biasCCP
classifier = BayesClassifier(priors, conds, eGain=gain)
# Classification should be 1,1,0. The update should alter, 0|1, 1|1, 2|0.
newCCP = biasCCP(classifier, .05)
# print(newCCP) # Uncomment to see if ccp comforms to predictiosn. Should bias 2|0 and raise 0 and 1 |1.
### Tests for biasMeasGenerator
generator = MeasurementGenerator((2, 2))
# print(generator.cmlProbs)
prev = generator.cmlProbs[1] - generator.cmlProbs[0]
# Let's give some conditionals that biases one measure
conds = ((.1, .7, .1, .1), (.25, .25, .25, .25))
# And feed biased priors
priors = (.7, .3)
biasMeasGenerator(generator, priors, conds)
# print(generator.cmlProbs)
now = generator.cmlProbs[1] - generator.cmlProbs[0]
# Should show bias towards second value now.
assert now > prev
generator = MeasurementGenerator((2, 2))
# print(generator.cmlProbs)
prev1 = generator.cmlProbs[1] - generator.cmlProbs[0]
prev2 = generator.cmlProbs[-1] - generator.cmlProbs[-2]
# Let's give some conditionals that biases one measure
conds = ((.1, .5, .1, .3), (.25, .25, .25, .25))
# And feed biased priors
priors = (.7, .3)
biasMeasGenerator(generator, priors, conds)
# print(generator.cmlProbs)
now1 = generator.cmlProbs[1] - generator.cmlProbs[0]
now2 = generator.cmlProbs[-1] - generator.cmlProbs[-2]
# Should show bias towards second value now.
assert now1 > prev1, (prev1, now1)
assert now2 > prev2, (prev2, now2)
def main():
priors = (.6, .4)
conds = [[.12 / .6, .18 / .6, .3 / .6], [.2 / .4, .16 / .4, .04 / .4]]
gain = ((1, 0), (0, 2))
### Tests for biasMeasGenerator
generator = MeasurementGenerator((2, 2))
# print(generator.cmlProbs) ## For peeking.
prev = generator.cmlProbs[1] - generator.cmlProbs[0]
# Let's give some conditionals that biases one measure
conds = ((.1, .7, .1, .1), (.25, .25, .25, .25))
# And feed biased priors
priors = (.7, .3)
biasMeasGenerator(generator, priors, conds)
# print(generator.cmlProbs)
now = generator.cmlProbs[1] - generator.cmlProbs[0]
# Should show bias towards second value now.
assert now > prev
generator = MeasurementGenerator((2, 2))
# print(generator.cmlProbs)
# Records previous probability for second value.
prev1 = generator.cmlProbs[1] - generator.cmlProbs[0]
# Records previous probability for final value.
prev2 = generator.cmlProbs[-1] - generator.cmlProbs[-2]
# Let's give some conditionals that biases one measure
conds = ((0.0, .5, 0.0, .5), (.05, .4, .05, .5))
# And feed biased priors
priors = (.7, .3)
biasMeasGenerator(generator, priors, conds)
# print(generator.cmlProbs)
now1 = generator.cmlProbs[1] - generator.cmlProbs[0]
now2 = generator.cmlProbs[-1] - generator.cmlProbs[-2]
# Should show bias towards second value now. May vary due to the original conditional probability being random.
assert now1 > prev1, (prev1, now1)
assert now2 > prev2, (prev2, now2)
def fitData(classifier: BayesClassifier, generator: MeasurementGenerator,
Z: int) -> Tuple[BayesClassifier, Tuple[int], Tuple[int]]:
# Uses classifier values to generate new measure and tag generators that are biased towards the classifier.
# Then generates Z new measure, tag pairs.
# Returns measure-tag pairs for another round of testing
conds = classifier.cond
priors = classifier.prior
posts = calcPosterior(priors, conds)
postsNorm = normPosterior(posts)
biasMeasGenerator(generator, priors, conds)
measures = generator.genMeas(Z)
tags = genBiasTags(measures, postsNorm)
return measures, tags
def fitData(classifier: BayesClassifier, generator: MeasurementGenerator,
Z: int,
delta: float) -> Tuple[BayesClassifier, Tuple[int], Tuple[int]]:
# Takes delta and uses to update conditional probability
# Uses new value to update distribution of measurement space
# Then generates Z new measures
# Then generates new tags in line with conditional probabbility distribution of class per tag
# Returns a classifier and measure-tag pairs for another round of testing
conds = classifier.cond
priors = classifier.prior
posts = calcPosterior(priors, conds)
postsNorm = normPosterior(posts)
biasMeasGenerator(generator, priors, conds)
measures = generator.genMeas(Z)
tags = genBiasTags(measures, postsNorm)
return classifier, measures, tags
def main():
parser = argparse.ArgumentParser()
parser.add_argument("samples",
help="Number of measurement samples to generate.",
type=int)
parser.add_argument("dimen", help="Measurement space.", type=str)
parser.add_argument("classes", help="Number of classes.", type=int)
parser.add_argument("seed",
help="Random seed for experiement duplication.",
type=int)
parser.add_argument(
"--vfolds",
"-v",
default=10,
help=
"Number of v-folds to partition testing data for v-folds testing. Default is 10.",
type=int)
parser.add_argument(
"--optimization",
"-o",
default=0.0,
help=
"Specify if iterative improvement of class conditional probability values should be taken.",
type=float)
parser.add_argument("--iteration",
"-t",
default=10,
help="Number of iterations for conditional update.",
type=int)
parser.add_argument(
"--identity",
"-i",
action="store_true",
default=False,
help="Specify if economic gain matrix should be identity.")
args = parser.parse_args()
# Checks that our given limits are even feasible memory wise
# Prompts for reader friendliness
print("Generating testing data for seed {}".format(args.seed))
# Sets seed
seed(args.seed)
# Assigns values
dimen = eval(args.dimen)
# Calculates size of domain
M = 1
for N in dimen:
M *= N
K = args.classes
V = args.vfolds
Z = args.samples
print("Dimensions of Measurement Space: {}".format(dimen))
print("Number of Samples: {}".format(Z))
print("Classes: {}".format(K))
# Checks that this is even possible to calculate.
if config.computeLimit(M, K):
print("Possible measurements exceed memory capabilities.")
sys.exit()
print("Generating {0}x{0} Gain Matrix. Identity Matrix: {1}".format(
K, args.identity))
gain = genGain(K, identity=args.identity)
print("{}x{} Economic Gain Matrix Generated".format(
len(gain), len(gain[0])))
# Generates measures
print("Generating {} Measure-Value pairs.".format(Z))
print("Generating measures.")
generator = MeasurementGenerator(dimen)
measures = generator.genMeas(Z)
assigner = ClassAssign(dimen, K)
tags = assigner.assign(measures)
print("{} measures and {} values generated.".format(
len(measures), len(tags)))
## Generates classifier.
print(
"Generating class conditional probabilities for {} classes and {} possible measures."
.format(K, M))
conditionals = genCCP(K, dimen)
print(
"Class conditional probabilities generated for {} classes and {} possible measures"
.format(len(conditionals), len(conditionals[0])))
classifier = BayesClassifier(
None, conditionals,
eGain=gain) # No priors given since vFold always assigns.
print("Testing classifier. V-fold factor: {}".format(V))
measures, tags = shuffle(measures, tags)
results = vFold(measures, tags, V, classifier)
matrix = genConMatrix(tags, results, K)
norm = normConMatrix(matrix)
expGain = calcExpGain(norm, classifier.eGain)
#expGain = test(classifier, measures, tags, V=V)
print("The expected gain for the given data is: {}".format(expGain))
#### Here we will work on updating
if args.optimization:
print(
"Fitting data for improved performance. Improvement factor {} used over {} iterations."
.format(args.optimization, args.iteration))
gains = []
# Going to set priors generated from this measurement set as permanent priors.
priors = calcClassProb(tags, K)
classifier.priorUpdate(priors)
for i in range(args.iteration):
# print(priors)
classifier.optimize(args.optimization, measures, tags)
classifier, measures, tags = fitData(classifier, generator, Z,
args.optimization)
measures, tags = shuffle(measures, tags)
results = vFold(measures, tags, V, classifier)
matrix = genConMatrix(tags, results, K)
norm = normConMatrix(matrix)
expGain = calcExpGain(norm, classifier.eGain)
#expGain = test(classifier, measures, tags, V=V)
gains.append(expGain)
print("Expected Gain from iteration {} is {}".format(
i + 1, expGain))
print("The expected gain for fitted data after {} iterations is: {}".
format(args.iteration, gains[-1]))
# Writes all data to files
print("Writing to file.")
reader.writeData(measures, tags, dimen)
reader.writePriors(classifier.prior)
reader.writeGain(gain)
reader.writeCCP(classifier.cond)
print("Done.")