# Check size
# print('* Params Size')
# print(mpl.size)
# print(t1.shape, t2.shape)
print('')
print('* Load debugging weights')
print(mpl.size)
# set initial weight
# mpl.params[0:(si + 1) * sh] = t1.reshape(((si + 1) * sh,))
# mpl.params[(si + 1) * sh:(sh + 1) * so + (si + 1) * sh] = t2.reshape((so * (sh + 1),))
# compute cost
print('* Cost Function: ' + str(mpl.cost_function(mpl.params)) + ' ' + str(2.10095))
print('Diff: ' + str(mpl.cost_function(mpl.params) - 2.10095))
print('')
# compute regularized cost
print('* Regularized Cost Function: ' + str(mpl.cost_function(mpl.params)) + ' ' + str(2.25231))
print('Diff: ' + str(mpl.cost_function(mpl.params) - 2.25231))
print('')
print('* Gradient Checking')
# g = mpl.usual_grad(mpl.params) # mpl.gradient(mpl.params, 0)
g = mpl.gradient(mpl.params)
num = mpl.numerical_gradient(mpl.params)
gr = mpl.gradient(mpl.params)
#0 0.3155 99.87 23.8 48.38 11.42
for s in struct:
Options["structure"]["hidden"] = s
Options['regularization'] = 1.0
info = str()
tp = Perceptron(train_set['y'], train_set['x'], Options)
print('\nStructure: ' + str(s) + '\tparams: ' + str(len(tp.params)))
print('idx cost acc acc norm time')
print('-----------------------------------------')
for i in range(0, test):
mpl = Perceptron(train_set['y'], train_set['x'], Options)
lb = lambda: mpl.lbfgs(ite_table[3])
time = tm.timeit(lb, number=1)
h1 = mpl.predict(train_set['x'], mpl.params)
h2 = mpl.predict(valid_set['x'], mpl.params)
info = str(i) + '\t' + \
str(mpl.cost_function(mpl.params))[0:6] + '\t' +\
str(mpl.accuracy(h1, train_set['yl'], 1))[0:5] + '\t' +\
str(mpl.accuracy(h2, valid_set['yl'], 1))[0:5] + '\t' +\
str(np.linalg.norm(mpl.params))[0:5] + '\t' +\
str(time)[0:5] + str('')
print(info)
