def parallel_probs(sim, A):
sim.initializeData(fit=None, full_H1=A)
fixes = sim.sample_many(num_samples = 2000)
sim2 = simulator.FixGen(fixes)
sim2.initializeData(fit=None)
stat_base = np.random.multinomial(sim2.nosamples[1],np.sum(sim.full_H1[1],1))
return (multinomial(stat_base,np.sum(sim.full_H1[1],1)), multinomial(np.sum(sim2.full_H1[1],1)*sim2.nosamples[1], np.sum(sim.full_H1[1],1)))
def test_sim(self):
### XXX: Look out for the first angles and lengths!
fm = fixmat.FixmatFactory('fixmat_photos.mat')
sim = simulator.FixGen(fm)
### Test case 1: uniform dist
H = np.ones((37,361))
H/=H.sum()
A = []
for i in range(36):
A.append(H)
sim.initializeData(fit=None, full_H1=A)
### XXX: Evtl. numsamples anpassen an die menge der benutzten bins.
fixes = sim.sample_many(num_samples = 500)
sim2 = simulator.FixGen(fixes)
sim2.initializeData(fit=None)
result_x = []
result_y = []
KLDiv_real_x = []
KLDiv_real_y = []
for i in range(len(sim.full_H1)-1):
KLDiv_base_x = []
KLDiv_base_y = []
tmp = []
tmp1 = []
prob_base_x = []
prob_base_y = []
prob_real_x = []
prob_real_y = []
### Draw 1000 samples from multinomial dist
stat_base_x = np.random.multinomial(sim2.nosamples[i],np.sum(sim.full_H1[i],1),size=500)
stat_base_y = np.random.multinomial(sim2.nosamples[i],np.sum(sim.full_H1[i],0),size=500)
# Compute probabilities of stat_base
for j in range(len(stat_base_x)):
tmp.append(multinomial(stat_base_x[j], np.sum(sim.full_H1[i],1)))
tmp1.append(multinomial(stat_base_y[j], np.sum(sim.full_H1[i],0)))
prob_base_x.append(tmp)
prob_base_y.append(tmp1)
### Compute KLDiv between 1000 samples and source
for j in range(len(stat_base_x)):
KLDiv_base_x.append(KLDiv(stat_base_x[j]/np.float(np.sum(stat_base_x[j])), np.sum(sim.full_H1[i],1)))
KLDiv_base_y.append(KLDiv(stat_base_y[j]/np.float(np.sum(stat_base_y[j])), np.sum(sim.full_H1[i],0)))
### Compute KLDiv between simulated and source
KLDiv_real_x.append(KLDiv(np.sum(sim.full_H1[i],1), np.sum(sim2.full_H1[i],1)))
KLDiv_real_y.append(KLDiv(np.sum(sim.full_H1[i],0), np.sum(sim2.full_H1[i],0)))
# Compute probabilities of real outcomes
prob_real_x.append(multinomial(np.sum(sim2.full_H1[i],1)*sim2.nosamples[i], np.sum(sim.full_H1[i],1)))
prob_real_y.append(multinomial(np.sum(sim2.full_H1[i],0)*sim2.nosamples[i], np.sum(sim.full_H1[i],0)))
### Compare simulated against multinomially sampled
result_x.append(np.array(KLDiv_real_x[-1]) > np.mean(KLDiv_base_x)+np.std(KLDiv_base_x))
result_y.append(np.array(KLDiv_real_y[-1]) > np.mean(KLDiv_base_y)+np.std(KLDiv_base_y))
pdb.set_trace()
KLDiv_real_x, KLDiv_real_y = np.array(KLDiv_real_x), np.array(KLDiv_real_y)
pdb.set_trace()
if sum(result_x)<len(result_x):
print "Horizontal margin failure in layer(s) " + repr(np.where(result_x))
print "With a KLDiv of " + repr(KLDiv_real_x[result_x])
if sum(result_y)<len(result_y):
print "Vertical margin failure in layer(s) " + repr(np.where(result_y))
print "With a KLDiv of " + repr(KLDiv_real_y[result_y])