# Create a new model, with this distribution
new_detector = detector.copy()
new_detector.kernel_templates = counts
new_detector.support = None
new_detector.use_alpha = False
# Return model
return new_detector
if __name__ == '__main__':
import argparse
from settings import load_settings
ag.set_verbose(True)
parser = argparse.ArgumentParser(
description="Convert model to integrate background model")
parser.add_argument('settings',
metavar='<settings file>',
type=argparse.FileType('r'),
help='Filename of settings file')
parser.add_argument('bkg',
metavar='<background file>',
type=argparse.FileType('rb'),
help='Background model file')
parser.add_argument('output',
metavar='<output file>',
type=argparse.FileType('wb'),
help='Model output file')
import numpy as np
import amitgroup as ag
import itertools as itr
import sys
import os
import pnet
import time
def test(ims, labels, net):
yhat = net.classify(ims)
return yhat == labels
if pnet.parallel.main(__name__):
ag.set_verbose(True)
print("1")
import argparse
parser = argparse.ArgumentParser()
#parser.add_argument('seed', metavar='<seed>', type=int, help='Random seed')
#parser.add_argument('param', metavar='<param>', type=string)
parser.add_argument('model',metavar='<model file>',type=argparse.FileType('rb'), help='Filename of model file')
print("ohhh")
parser.add_argument('data',metavar='<mnist data file>',type=argparse.FileType('rb'),help='Filename of data file')
parser.add_argument('label',metavar='<mnist data file>',type=argparse.FileType('rb'),help='Filename of data file')
parser.add_argument('numOfClassModel',metavar='<numOfClassModel>', type=int, help='num Of Class Model')
args = parser.parse_args()
param = args.model
def trial(seed=0):
eps = 1e-10
start = time()
ag.set_verbose(True)
np.set_printoptions(precision=2, suppress=True)
K = 2
eachN = 100000
N = K * eachN
M = 500
np.random.seed(0)
theta = np.random.random((K, M))
alphas = (np.random.random((K, M)) < 0.5).astype(float)
print alphas.mean()
#alphas = np.tile(np.random.random(M) < 0.5, (K, 1))
#b = np.clip(np.random.random(M), 0.1, 0.9)
b = np.ones(M) * 0.2
print theta
print b
np.random.seed(seed + 1000)
X = np.zeros((N, M))
A = np.zeros((N, M))
for k in xrange(K):
for i in xrange(eachN):
X[k * eachN + i] = np.random.random(M) < theta[k]
A[k * eachN + i] = alphas[k]
#X = (np.random.random((N, M)) < 0.5).astype(float)
X *= A
end = time()
print end - start
model = ag.stats.BernoulliMixture(K, X.astype(np.uint8))
model.run_EM(eps, 1e-5)
support = model.remix(A)
corrected = model.templates + (1 - support) * b
print "X"
print X
print "alphas"
print A
#print "b"
#print b
N = K * eachN
X2 = np.zeros((N, M))
A2 = np.zeros((N, M))
for k in xrange(K):
for i in xrange(eachN):
back = (np.random.random(M) < b).astype(float)
fore = (np.random.random(M) < theta[k]).astype(float)
X2[k * eachN + i] = alphas[k] * fore + (1 - alphas[k]) * back
model2 = ag.stats.BernoulliMixture(K, X2.astype(np.uint8))
model2.run_EM(eps, 1e-5)
#print X
#print X2
aff1 = model.affinities
aff2 = model2.affinities
print 's', aff1.shape
print 'support'
print support
model2templates = np.clip(model2.templates, 0.05, 0.95)
corrected = np.clip(corrected, 0.05, 0.95)
model1templates = np.clip(model.templates, 0.05, 0.95)
both = np.fabs(
[model2templates - corrected,
model2templates[::-1] - corrected]).sum(axis=1).sum(axis=1)
if np.argmin(both) == 1:
model2templates = model2templates[::-1]
aff2 = aff2[:, ::-1]
#print aff1
#print aff2
print 'model'
print model1templates
print 'corrected'
print corrected
print 'model2'
print model2templates
print '---'
scores = np.fabs(corrected -
model2templates).sum(), np.fabs(model1templates -
model2templates).sum()
print scores[0], scores[1]
return corrected, model2templates, scores
alpha = args.alpha[0]
bfrom, bto, eta, bN = args.test_conjugate
bN = int(bN)
surplus_file = args.surplus[0]
eta0 = args.penalty[0]
rho0 = args.rho[0]
if deform_type == 'none':
deform_type = None
verbose = args.verbose
import numpy as np
from classifier import classify, surplus
from scipy.optimize import fmin
import amitgroup as ag
ag.set_verbose(verbose)
features_data = np.load(feat_file)
all_features = features_data['features']
all_labels = features_data['labels']
mixtures_data = np.load(mixtures_file)
all_templates = mixtures_data['templates']
coefs = np.load(coef_file)
try:
all_graylevels = features_data['originals']
all_graylevel_templates = mixtures_data['graylevel_templates']
except KeyError:
raise Exception("The feature file must be run with --save-originals, and the mixtures must be trained with this file")
alpha = args.alpha[0]
bfrom, bto, eta, bN = args.test_conjugate
bN = int(bN)
surplus_file = args.surplus[0]
eta0 = args.penalty[0]
rho0 = args.rho[0]
if deform_type == 'none':
deform_type = None
verbose = args.verbose
import numpy as np
from classifier import classify, surplus
from scipy.optimize import fmin
import amitgroup as ag
ag.set_verbose(verbose)
features_data = np.load(feat_file)
all_features = features_data['features']
all_labels = features_data['labels']
mixtures_data = np.load(mixtures_file)
all_templates = mixtures_data['templates']
coefs = np.load(coef_file)
try:
all_graylevels = features_data['originals']
all_graylevel_templates = mixtures_data['graylevel_templates']
except KeyError:
raise Exception(
"The feature file must be run with --save-originals, and the mixtures must be trained with this file"
)
def trial(seed=0):
eps = 1e-10
start = time()
ag.set_verbose(True)
np.set_printoptions(precision=2, suppress=True)
K = 2
eachN = 100000
N = K*eachN
M = 500
np.random.seed(0)
theta = np.random.random((K, M))
alphas = (np.random.random((K, M)) < 0.5).astype(float)
print alphas.mean()
#alphas = np.tile(np.random.random(M) < 0.5, (K, 1))
#b = np.clip(np.random.random(M), 0.1, 0.9)
b = np.ones(M) * 0.2
print theta
print b
np.random.seed(seed+1000)
X = np.zeros((N, M))
A = np.zeros((N, M))
for k in xrange(K):
for i in xrange(eachN):
X[k*eachN+i] = np.random.random(M) < theta[k]
A[k*eachN+i] = alphas[k]
#X = (np.random.random((N, M)) < 0.5).astype(float)
X *= A
end = time()
print end - start
model = ag.stats.BernoulliMixture(K, X.astype(np.uint8))
model.run_EM(eps, 1e-5)
support = model.remix(A)
corrected = model.templates + (1-support) * b
print "X"
print X
print "alphas"
print A
#print "b"
#print b
N = K*eachN
X2 = np.zeros((N, M))
A2 = np.zeros((N, M))
for k in xrange(K):
for i in xrange(eachN):
back = (np.random.random(M) < b).astype(float)
fore = (np.random.random(M) < theta[k]).astype(float)
X2[k*eachN+i] = alphas[k] * fore + (1-alphas[k]) * back
model2 = ag.stats.BernoulliMixture(K, X2.astype(np.uint8))
model2.run_EM(eps, 1e-5)
#print X
#print X2
aff1 = model.affinities
aff2 = model2.affinities
print 's', aff1.shape
print 'support'
print support
model2templates = np.clip(model2.templates, 0.05, 0.95)
corrected = np.clip(corrected, 0.05, 0.95)
model1templates = np.clip(model.templates, 0.05, 0.95)
both = np.fabs([model2templates - corrected, model2templates[::-1] - corrected]).sum(axis=1).sum(axis=1)
if np.argmin(both) == 1:
model2templates = model2templates[::-1]
aff2 = aff2[:,::-1]
#print aff1
#print aff2
print 'model'
print model1templates
print 'corrected'
print corrected
print 'model2'
print model2templates
print '---'
scores = np.fabs(corrected - model2templates).sum(), np.fabs(model1templates - model2templates).sum()
print scores[0], scores[1]
return corrected, model2templates, scores
def trial(seed=0):
eps = 1e-10
start = time()
ag.set_verbose(True)
np.set_printoptions(precision=2, suppress=True)
K = 2
eachN = 100000
N = K * eachN
M = 40
pool_length = 2
np.random.seed(0)
theta = np.random.random((K, M))
alphas = (np.random.random((K, M)) < 0.5).astype(float)
print alphas.mean()
# alphas = np.tile(np.random.random(M) < 0.5, (K, 1))
# b = np.clip(np.random.random(M), 0.1, 0.9)
b = np.ones(M) * 0.2
b2 = np.ones(M // pool_length) * 0.36
print theta
print b
np.random.seed(seed + 1000)
X = np.zeros((N, M))
A = np.zeros((N, M))
for k in xrange(K):
for i in xrange(eachN):
X[k * eachN + i] = np.random.random(M) < theta[k]
A[k * eachN + i] = alphas[k]
# X = (np.random.random((N, M)) < 0.5).astype(float)
X *= A
end = time()
print end - start
Y = maxpool(X, pool_length)
model = ag.stats.BernoulliMixture(K, Y.astype(np.uint8))
model.run_EM(eps, 1e-5)
B = meanpool(A, pool_length)
B2 = maxpool(A, pool_length)
print "#" * 80
print A.shape
print B.shape
support = model.remix(B)
support2 = model.remix(B2)
# corrected = model.templates + (1-(support+support2)/2) * b2
corrected = 1 - (1 - model.templates) * (1 - b[0]) ** (pool_length * (1 - support))
print "X"
print X
print "alphas"
print A
# print "b"
# print b
N = K * eachN
X2 = np.zeros((N, M))
A2 = np.zeros((N, M))
for k in xrange(K):
for i in xrange(eachN):
back = (np.random.random(M) < b).astype(float)
fore = (np.random.random(M) < theta[k]).astype(float)
X2[k * eachN + i] = alphas[k] * fore + (1 - alphas[k]) * back
print "---------->"
print X2.shape
Y2 = maxpool(X2, pool_length)
print Y2.shape
model2 = ag.stats.BernoulliMixture(K, Y2.astype(np.uint8))
model2.run_EM(eps, 1e-5)
# print X
# print X2
aff1 = model.affinities
aff2 = model2.affinities
print "s", aff1.shape
print "support"
print support
model2templates = np.clip(model2.templates, 0.05, 0.95)
corrected = np.clip(corrected, 0.05, 0.95)
model1templates = np.clip(model.templates, 0.05, 0.95)
both = np.fabs([model2templates - corrected, model2templates[::-1] - corrected]).sum(axis=1).sum(axis=1)
if np.argmin(both) == 1:
model2templates = model2templates[::-1]
aff2 = aff2[:, ::-1]
# print aff1
# print aff2
print "model"
print model1templates
print "corrected"
print corrected
print "model2"
print model2templates
print "---"
print (corrected - model2templates)
scores = np.fabs(corrected - model2templates).sum(), np.fabs(model1templates - model2templates).sum()
print scores[0], scores[1]
return corrected, model2templates, scores
