def compare(method_1="bsg", method_2="nb", dataset=4):
# load dataset
n = 200
if dataset == 4:
x, y = load4(n)
elif dataset == 5:
x, y = load5(n)
# optimize using two methods
if method_1 == "bsg":
fun1 = bsg.optimize(x, y)
label1 = "steepest descent"
elif method_1 == "nb":
fun1 = nb.optimize(x, y)
label1 = "Newton"
elif method_1 == "mn":
fun1 = mn.optimize(x, y)
label1 = "Newton"
elif method_1 == "mssg":
fun1 = mssg.optimize(x, y)
label1 = "simple steepest descent"
if method_2 == "bsg":
fun2 = bsg.optimize(x, y)
label2 = "steepest descent"
elif method_2 == "nb":
fun2 = nb.optimize(x, y)
label2 = "Newton"
elif method_2 == "mn":
fun2 = mn.optimize(x, y)
label2 = "Newton"
elif method_2 == "mssg":
fun2 = mssg.optimize(x, y)
label2 = "simple steepest descent"
diff_1 = []
diff_2 = []
for i in range(len(fun1) - 1):
diff_1.append(fun1[i] - fun1[-1])
for i in range(len(fun2) - 1):
diff_2.append(fun2[i] - fun2[-1])
plt.yscale('log')
plt.plot(np.arange(0, len(diff_1), 1),
diff_1,
marker="o",
color="red",
linewidth=0.5,
linestyle="--",
markersize=1,
label=label1)
plt.plot(np.arange(0, len(diff_2), 1),
diff_2,
marker="v",
color="blue",
linewidth=0.5,
linestyle=":",
markersize=1,
label=label2)
plt.ylabel("$J(w^{(t)}) - J(\hat{w})$")
plt.xlabel("iteration")
#plt.yscale('log')
plt.xlim(0, max(len(diff_1), len(diff_2)))
#plt.ylim(0, 0.1)
plt.legend()
path = "results/"
try:
os.mkdir(path)
except OSError:
print("Creation of the directory %s failed or Already exists" % path)
else:
print("Successfully created the directory %s " % path)
plt.savefig(path + method_1 + method_2 + ".png")
plt.savefig(path + method_1 + method_2 + ".eps")
plt.close()
return
softmax = exps / np.sum(exps)
#print(softmax)
for j in range(class_num):
if (j == yi):
grad[:,yi] += ((softmax[yi] - 1).reshape((1,1)) * xi).reshape((5,1,1))
else:
grad[:,j] += (softmax[j].reshape((1,1)) * xi).reshape((5,))
L -= (np.log(softmax[yi]))
#print(grad)
grad /= n
grad += 2 * lam * w
L /= n
obj_fun += L
print(str(step).rjust(4) + ": " + str(np.linalg.norm(grad)).ljust(22) + " " + str(np.asscalar(obj_fun)))
obj_fun_history.append(np.asscalar(obj_fun))
if (np.linalg.norm(grad) < tol):
break
w -= alpha * grad
step += 1
print("---------------------------------------------------------------")
print("step: " + str(step) + ", gnorm: " + str(np.linalg.norm(grad)) + ", obj_fun: " + str(np.asscalar(obj_fun)))
print("Omptimization was done.")
return obj_fun_history
if __name__ == '__main__':
x, y = load5(200)
optimize(x, y)
print('module name：{}'.format(__name__))