In [1]:
%pylab inline
In [2]:
import struct
In [3]:
!cd /tmp/; wget http://razik.univ-tln.fr/misc/I322/I322_tp.tgz ; tar xzvf /tmp/I322_tp.tgz
In [16]:
# f_name = "/home/partage/I322/RAW/a.raw"
f_name = "/tmp/a.raw"
In [5]:
def lire_son(f_name):
valeurs = []
with open(f_name, 'br') as fichier:
while fichier:
v = fichier.read(2)
if not v:
break
valeurs.append(struct.unpack('h', v)[0])
return(array(valeurs))
In [17]:
valeurs = lire_son(f_name)
In [18]:
#f = figure(figsize=(12,5))
f, ax = subplots()
figsize(12,5)
ax.set_axis_off()
plot(valeurs)
Out[18]:
In [8]:
def zcr(fenetre):
cpt = 0
old_val = fenetre[0]
for v in fenetre:
if v*old_val < 0:
cpt += 1
old_val = v
return cpt
In [76]:
zcr(valeurs)
Out[76]:
In [10]:
len(valeurs)
Out[10]:
In [11]:
range(220, len(valeurs), 220)
Out[11]:
In [100]:
les_zcr = [zcr(valeurs[deb:deb+220]) for deb in range(0, len(valeurs), 220)]
In [13]:
plot(les_zcr)
Out[13]:
In [14]:
plot(les_zcr)
Out[14]:
In [15]:
hist(les_zcr)
Out[15]:
In [16]:
hist(les_zcr)
Out[16]:
In [17]:
mean(les_zcr)
Out[17]:
In [18]:
std(les_zcr)
Out[18]:
In [19]:
def gauss(x, mu, sigma):
return 1/(sqrt(2*pi)*sigma) * exp(-0.5*(x - mu)**2/sigma**2)
In [20]:
Y = [gauss(x, 49.3, 20.22) for x in range(101)]
In [21]:
plot(Y)
plot(les_zcr, [0]*len(les_zcr), 'ro')
Out[21]:
In [22]:
from sklearn import mixture
In [103]:
g = mixture.GaussianMixture(1, covariance_type='diag')
In [24]:
g.fit(array(les_zcr).reshape(-1,1))
Out[24]:
In [25]:
g.means_
Out[25]:
In [26]:
g.covariances_
Out[26]:
In [27]:
g.weights_
Out[27]:
In [28]:
Z = [exp(g.score(x.reshape(-1,1))) for x in arange(101)]
In [29]:
plot(Z)
Out[29]:
In [104]:
a_g2 = mixture.GaussianMixture(2, covariance_type='diag')
In [45]:
a_g2.fit(array(les_zcr).reshape(-1,1))
Out[45]:
In [46]:
a_g2.means_
Out[46]:
In [47]:
a_g2.covariances_
Out[47]:
In [48]:
a_g2.weights_
Out[48]:
In [49]:
Z = [exp(a_g2.score(x.reshape(-1,1))) for x in arange(101)]
In [50]:
plot(Z)
Out[50]:
In [51]:
test_f_name = "/tmp/I322/RAW/0.raw"
In [52]:
test_valeurs = lire_son(test_f_name)
In [53]:
test_les_zcr = [zcr(test_valeurs[deb:deb+220]) for deb in range(0, len(test_valeurs), 220)]
In [54]:
plot(test_les_zcr)
Out[54]:
In [56]:
len(test_les_zcr)
Out[56]:
In [55]:
hist(test_les_zcr)
Out[55]:
In [58]:
a_g2.score(array([test_les_zcr[0]]).reshape(-1,1))
Out[58]:
In [61]:
test_les_zcr[0]
Out[61]:
In [65]:
def vraisemblance(x):
return 0.21988493*gauss(x, 79.11111498 , sqrt(10.77008664)) + 0.78011507*gauss(x, 40.90125872, sqrt(200.13660043))
In [66]:
# vraisemblance pour la première valeur des zcr
# c_g1*gauss(x, mu_g1, sigma_g1) + c_g2*gauss(x, mu_g2, sigma_g2)
# v = 0.21988493*gauss(0, 79.11111498 , sqrt(10.77008664)) + 0.78011507*gauss(0, 40.90125872, sqrt(200.13660043))
v = vraisemblance(test_les_zcr[0])
print(v)
print(log(v))
In [67]:
P = prod(array([vraisemblance(x) for x in test_les_zcr]))
print(P)
In [69]:
log(P)
Out[69]:
In [70]:
P_log = sum(array([log(vraisemblance(x)) for x in test_les_zcr]))
print(P_log)
In [72]:
a_g2.score(array(test_les_zcr).reshape(-1,1))
Out[72]:
In [79]:
P_a = prod(array([vraisemblance(x) for x in les_zcr]))
print(log(P_a))
In [105]:
zero_g2 = mixture.GaussianMixture(2, covariance_type='diag')
In [86]:
zero_g2.fit(array(les_zcr).reshape(-1,1))
Out[86]:
In [88]:
zero_g2.means_
Out[88]:
In [89]:
zero_g2.covariances_
Out[89]:
In [90]:
zero_g2.weights_
Out[90]:
In [93]:
Z = [exp(zero_g2.score(x.reshape(-1,1))) for x in arange(101)]
In [94]:
plot(Z)
Out[94]:
In [95]:
def vraisemblance_0(x):
return 0.53514291*gauss(x, 31.21117743 , sqrt(6.64596133)) + 0.46485709*gauss(x, 29.43841612, sqrt(344.80709667))
In [102]:
P_0 = prod(array([vraisemblance_0(x) for x in test_les_zcr]))
print(log(P_0))
In [101]:
P_0 = prod(array([vraisemblance_0(x) for x in les_zcr]))
print(log(P_0))
In [ ]:
# ce qui reste à faire
# 1. faire un modèle pour chaque classe (a ... z, 0 ... 9)
# 2. tester chaque fichier son sur l'ensemble des modèles (prise de décision)
# 3. construire la matrice de confusion (à partir des résultats précédents)