def choose_lambda(self):
'''train with different regularization parameters and choose
the one that minimizes the cost in the evaluation set.'''
tinit = 0.005* np.random.rand(self.LABS, self.N)
# initialize some working vars
#rango = np.array([1e-3, 1e-2, 1e-1, 1, 10, 100])
rango = np.array([2, 3, 4, 5, 6, 7, 8, 9, 10])
Jt, Je = np.array([]), np.array([])
bestC, bestL = 1e+10, 0.0
# cycle through lambdas and choose the one with lowest cost
for chosen_lambda in rango:
t = soft.optimizeThetas(tinit, self.xt, self.gt, \
numLabels=self.LABS, l=chosen_lambda, visual=True)
cost_t = soft.j(t, self.xt, self.gt, self.LABS, chosen_lambda)
cost_e = soft.j(t, self.xe, self.ge, self.LABS, chosen_lambda)
Jt = np.append(Jt, cost_t)
Je = np.append(Je, cost_e)
print 'chosen_lambda:', chosen_lambda
if cost_e < bestC:
bestC = cost_e
bestL = chosen_lambda
print '_______________new best is', bestL, 'with cost_e', cost_e
print "\n\nthe best lambda is", bestL
# plot
#line1 = plt.plot(np.log10(rango), Jt)
#line2 = plt.plot(np.log10(rango), Je)
line1 = plt.plot(rango, Jt)
line2 = plt.plot(rango, Je)
plt.setp(line1, linewidth=2.0, label='training', color='b', solid_joinstyle='round')
plt.setp(line2, linewidth=2.0, label='training', color='r', solid_joinstyle='round')
plt.xlabel('Lambda')
plt.ylabel('J')
plt.show()
pass
def learning_curves(self):
tinit = 0.005* np.random.rand(self.LABS, self.N)
m, n = self.xt.shape
sample = np.array([3, 6, 9, 12, 15, 18, 21, 24, 27, 30])*1000
Jt, Je = np.array([]), np.array([])
for m in sample:
my_t = soft.optimizeThetas(tinit, self.xt[0:m,:], self.gt[0:m,:], \
numLabels=self.LABS, l=self.L, visual=False)
Jt = np.append(Jt, soft.j(my_t, self.xt[0:m,:], self.gt[0:m,:], self.LABS, self.L))
Je = np.append(Je, soft.j(my_t, self.xe, self.ge, self.LABS, self.L))
# plot (m, Jtr) and (m, Jcv)
line1 = plt.plot(sample, Jt)
line2 = plt.plot(sample, Je)
plt.setp(line1, linewidth=2.0, label='training', color='b', solid_joinstyle='round')
plt.setp(line2, linewidth=2.0, label='training', color='r', solid_joinstyle='round')
plt.xlabel('Number of Examples')
plt.ylabel('Cost / Error')
plt.show()
pass