def Find_al_max_on_05(self, r_0, r_1):
al_max = 0.5
r_2 = self.Lin_W_on_al(al_max, r_0, r_1)
self.From_W_arr_to_Net(r_2)
r_2["iq"], iq_Test_min, iq_Test_max = iq_test(self.net, X, Y)
i_05 = r_2["iq"]
aa = 2 * (2 * r_1["iq"] + 2 * r_0["iq"] - 4 * i_05)
if aa < 0:
al_max = (r_1["iq"] + 3 * r_0["iq"] - 4 * i_05) / aa
r_2 = self.Lin_W_on_al(al_max, r_0, r_1)
self.From_W_arr_to_Net(r_2)
r_2["iq"], iq_Test_min, iq_Test_max = iq_test(self.net, X, Y)
return al_max, r_2
def rec_test(self, rec_tmp):
self.From_W_arr_to_Net(rec_tmp)
rec_tmp['iq'], iq_Test_min, iq_Test_max = iq_test(
self.net, self.X, self.Y)
rec_tmp["norm"] = self.calc_norm_rec(rec_tmp)
if rec_tmp["name"][0] == 'M':
rec_tmp["name"] = 'C' + rec_tmp["name"]
return rec_tmp
def Move_min_max_in_iq(self, tmp_matrix, step_min):
for i in range(len(tmp_matrix)):
if i != i_max:
tmp_matrix[i]["iq_del"] = (tmp_matrix[i_max]['iq'] -
tmp_matrix[i]['iq'])
r1 = self.Rec_from_to(tmp_matrix[i], tmp_matrix[i_max])
tmp_matrix[i]["step_max"] = self.calc_norm_rec(r1)
if tmp_matrix[i]["step_max"] < 0.1 * step_min:
r1 = self.Net_Grad_to_rec(r1)
tmp_matrix[i] = self.Move_Rec_Point_on_M(
tmp_matrix[i], r1, step_min)
tmp_matrix[i]["step_max"] = self.calc_norm_rec(
self.Rec_from_to(tmp_matrix[i], tmp_matrix[i_max]))
tmp_matrix[i]['iq'], iq_Test_min, iq_Test_max = iq_test(
self.net, X, Y)
tmp_matrix[i]["iq_del"] = (tmp_matrix[i_max]['iq'] -
tmp_matrix[i]['iq'])
if tmp_matrix[i]["step_max"] < step_min:
tmp_matrix[i] = self.Move_Rec_Point_on_M(
tmp_matrix[i], r1, 3 * tmp_matrix[i]["step_max"])
tmp_matrix[i]["step_max"] = self.calc_norm_rec(
self.Rec_from_to(tmp_matrix[i], tmp_matrix[i_max]))
tmp_matrix[i]['iq'], iq_Test_min, iq_Test_max = iq_test(
self.net, X, Y)
tmp_matrix[i]["iq_del"] = (tmp_matrix[i_max]['iq'] -
tmp_matrix[i]['iq'])
if tmp_matrix[i]['iq'] < iq_min:
i_min = i
iq_min = tmp_matrix[i]['iq']
if tmp_matrix[i]["step_max"] > 0:
i = i
else:
tmp_matrix[i]["step_max"] = 1000
tmp_matrix[i]['iq'] = tmp_matrix[i_min]['iq']
tmp_matrix[i]["iq_del"] = tmp_matrix[i_min]['iq_del']
else:
tmp_matrix[i]["step_max"] = 0
tmp_matrix[i]["iq_del"] = 0
return i_min, i_max
def Find_max_on_2(self, r_0, r_1):
r_2 = self.Lin_W_on_al(2, r_0, r_1)
self.From_W_arr_to_Net(r_2)
r_2["iq"], iq_Test_min, iq_Test_max = iq_test(self.net, X, Y)
iq_2 = r_2["iq"]
a = (r_0["iq"] + r_2["iq"] - 2 * r_1["iq"]) / 2
b = r_1["iq"] - r_0["iq"] - a
al_max = -b / (2 * a)
extr = a * al_max * al_max + b * al_max + r_0["iq"]
if ((r_2["iq"] > r_1["iq"]) and (r_2["iq"] > r_0["iq"])
and (extr < r_2["iq"])):
al_max = 2
print(' ', r_2["name"], '->', '{0:.4f}'.format(r_0["iq"]),
'{0:.4f}'.format(r_1["iq"]), '{0:.4f}'.format(iq_2),
'---------->', '{0:.4f}'.format(r_2["iq"]))
return r_2
if ((extr < r_1["iq"]) and (extr < r_0["iq"])):
if (r_0["iq"] < r_1["iq"]):
al_max = 1.1
else:
al_max = -0.1
if al_max > 15:
al_max = 15
if al_max < -2:
al_max = -2
al_max
r_2 = Lin_W_on_al(al_max, r_0, r_1)
self.From_W_arr_to_Net(r_2)
r_2["iq"], iq_Test_min, iq_Test_max = iq_test(self.net, X, Y)
print(' M_al :', r_2["name"], 'EXTR_2=', '{0:.3f}'.format(extr),
'->', '{0:.3f}'.format(r_0["iq"]), '{0:.3f}'.format(r_1["iq"]),
'{0:.3f}'.format(iq_2), '---------->',
'{0:.3f}'.format(r_2["iq"]))
return r_2
def Move_Rec_Point_on_M(self, Rec_P, Rec_M, KoefStep):
R_1 = copy.deepcopy(Rec_P)
for i in range(len(Rec_P["weights"])):
for j in range(len(Rec_P["weights"][i])):
for k in range(len(Rec_P["weights"][i][j])):
aa = KoefStep * Rec_M["weights"][i][j][k]
R_1["weights"][i][j][k] += aa
self.From_W_arr_to_Net(R_1)
R_1["iq"], iq_Test_min, iq_Test_max = iq_test(self.net, X, Y)
R_1["name"] = Rec_P["name"]
if R_1["name"][0] != 'M':
R_1["name"] = 'M_' + R_1["name"]
return R_1
' ---- n_GR = ' + '{0:.2f}'.format(net.norm_G) +
' -----Learn_rate = ' + '{0:.4f}'.format(l_r))
N_min, N_max = f.find_i_min_max_in_iq(W_array, N_min, N_max)
f.print_w_arr(
W_array, ' i = ' + str(i_epoch) + ' ii=' + str(ii_epoch) +
' ---2 n_GR = ' + '{0:.2f}'.format(net.norm_G) +
' -----Learn_rate = ' + '{0:.4f}'.format(l_r))
IQ_in = W_array[N_max]['iq']
f.X, f.Y, f.name_of_answer_array = f.sorted_Test_data(
f.X, f.Y, f.name_of_answer_array)
#con = input('Press any key...')
net.iq, net.iq_min, net.iq_max = iq_test(f.net, f.X, f.Y)
teacher.train(
f.X,
f.Y,
25, #learning_rate = [l_r, l_r_old],
name_of_answer=name_of_answer_array,
random_data=True,
save_output_data=True)
net.iq, net.iq_min, net.iq_max = f.iq_test(net, X, Y)
R_train = f.Net_to_W_arr(W_array[N_max], 'tr_' + str(net.total_epoch))
W_array.append(R_train)
print(i_epoch, ii_epoch, " ---- End train -- ", R_train['name'],
def Find_max_on_Line(self, r_1, r_2):
delta_al = 0.5
almax_old = 0
delta_s = 1000
nn = 0
if r_1['iq'] < r_2['iq']:
r_min = r_1
r_max = copy.deepcopy(r_2)
else:
r_min = r_2
r_max = copy.deepcopy(r_1)
while ((delta_al > 0.01) and (nn < 10) and delta_s > 1):
nn += 1
r_1_2 = self.Rec_from_to(r_min, r_max)
delta_s = r_1_2['norm']
almax, r_3 = self.Find_al_max_on_05(r_min, r_max)
delta_al = math.fabs(r_3['iq'] - r_max['iq'])
iq_min = r_min['iq']
iq_max = r_max['iq']
iq_new = r_3['iq']
#print('{0:.5f}'.format(r_max['iq']),'{0:.5f}'.format(r_3['iq']), '{0:.5f}'.format(delta_al))
if (r_3['iq'] > r_max['iq']):
r_min = r_max
r_max = copy.deepcopy(r_3)
str_ = 'var_Best'
else:
str_ = 'var_0.95'
almax = 0.95
r_3 = self.Lin_W_on_al(0.95, r_min, r_max)
self.From_W_arr_to_Net(r_3)
r_3["iq"], iq_Test_min, iq_Test_max = iq_test(self.net, X, Y)
iq_new = r_3["iq"]
delta_al = math.fabs(r_3['iq'] - r_max['iq'])
if (r_3['iq'] > r_max['iq']):
str_ = 'yes_0.95'
r_min = r_max
r_max = copy.deepcopy(r_3)
else:
r_min = r_3
str_ = 'var_s_so'
print(
' Find_on_Line n=',
nn,
str_,
"\t",
'{0:.4f}'.format(iq_min),
'{0:.4f}'.format(iq_max),
'{0:.4f}'.format(iq_new),
'<--al=',
'{0:.2f}'.format(almax),
' delta=',
'{0:.4f}'.format(delta_al),
' del_s=',
'{0:.1f}'.format(delta_s),
)
r_max['name'] = 'FL_' + str(self.net.total_epoch)
r_1_2 = self.Rec_from_to(r_min, r_max)
#From_W_arr_to_Net(r_max)
return nn, r_max, r_1_2
err_min = err_ - 3*sigma
if err_max < 0:
err_max = 0
iq_min = (1 - err_max)*100
iq_max = (1 - err_min)*100
iq = (iq_min + iq_max)/2
logging.info("%s | iq_test | iq: %s | iq_min: %s | iq_max: %s | nn_IQ: %s" %
(datetime.datetime.today().strftime("%Y-%m-%d-%H.%M.%S"),
iq, iq_min, iq_max, len(X)))
return iq, iq_min, iq_max'''
net.load()
Y_ = []
X_ = []
for i in range(0, 1500):
X, Y = random_()
# Y = Image.open("canon/can_"+str(i)+".png").convert("L")
# Y = np.array(Y)
# Y = Y.flatten()/255
Y_.append(Y)
X_.append(net.forward(Y, training=True))
print(iq_test(net, X_, Y_))
from netpy.nets import FeedForwardNet
from netpy.tools.functions import iq_test
import numpy as np
net = FeedForwardNet(name='N_al_gr')
net.load_net_data()
net.load_weights()
test_X = np.loadtxt('data_set/test_X.txt')
test_Y = np.loadtxt('data_set/test_Y.txt')
print(iq_test(net, test_X, test_Y))