def losc_test(paramlist, show, val):
fimtgd = FIMTGD(gamma=paramlist[0],
n_min=paramlist[1],
alpha=paramlist[2],
threshold=paramlist[3],
learn=paramlist[4])
fimtls = FIMTLS(gamma=paramlist[0],
n_min=paramlist[1],
alpha=paramlist[2],
threshold=paramlist[3],
learn=paramlist[4])
gfimtls = gFIMTLS(gamma=paramlist[0],
n_min=paramlist[1],
alpha=paramlist[2],
threshold=paramlist[3],
learn=paramlist[5])
cumLossgd = [0]
cumLossls = [0]
cumLossgls = [0]
data = generate_Losc(4000)
data = np.array(sorted(data, key=lambda x: x[0]))
o_target = data[:, -1]
input = data[:, 1:-1]
for counter in range(len(data)):
noise = (np.random.uniform() - 0.5) * 0.8
target = o_target[counter] + noise
cumLossgd.append(
cumLossgd[-1] +
np.fabs(o_target[counter] -
fimtgd.eval_and_learn(np.array(input[counter]), target)))
cumLossls.append(
cumLossls[-1] +
np.fabs(o_target[counter] -
fimtls.eval_and_learn(np.array(input[counter]), target)))
cumLossgls.append(
cumLossgls[-1] +
np.fabs(o_target[counter] -
gfimtls.eval_and_learn(np.array(input[counter]), target)))
if show:
f = plt.figure()
plt.plot(cumLossgd[1:], label="Gradient Descent Loss")
f.hold(True)
plt.plot(cumLossls[1:], label="Filter Loss")
# avglossgd=np.array([cumLossgd[-1]/len(cumLossgd)]*len(cumLossgd))
# plt.plot(avglossgd,label="Average GD Loss")
# plt.plot([cumLossls[-1]/len(cumLossls)]*len(cumLossls), label="Average Filter Loss")
plt.title("CumLoss Ratio:" + str(
min(cumLossgd[-1], cumLossls[-1]) /
max(cumLossgd[-1], cumLossls[-1])))
plt.legend()
figname = "g" + str(paramlist[0]) + "_nmin" + str(paramlist[1]) + "_al" + str(paramlist[2]) + "_thr" + str(
paramlist[3]) \
+ "_lr" + str(paramlist[4]) + ".png"
plt.savefig(figname)
# plt.show()
f.clear()
return [cumLossgd, cumLossls, cumLossgls, val, paramlist]
def sine_test(paramlist,show,val):
#print(val)
#print(paramlist)
fimtgd=FIMTGD(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[4])
fimtls=FIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[4])
gfimtls=gFIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[5])
cumLossgd =[0]
cumLossls =[0]
cumLossgls =[0]
if True:
start = 0.0
end = 1.0
x = list()
y = list()
for i in range(4000):
input = np.random.uniform(0.0,1.0)*2*np.pi
target = np.sin(input)
if i > 2000:
target += 1.0
o_target = target
noise = (np.random.uniform() - 0.5) * 0.8
target += noise
x.append(input)
y.append(target)
cumLossgd.append(cumLossgd[-1] + np.sqrt(np.fabs(o_target - fimtgd.eval_and_learn(np.array(input), target))**2))
cumLossls.append(cumLossls[-1] + np.sqrt(np.fabs(o_target - fimtls.eval_and_learn(np.array(input), target))**2))
cumLossgls.append(cumLossgls[-1] + np.sqrt(np.fabs(o_target - gfimtls.eval_and_learn(np.array(input), target))**2))
#plt.scatter(x=x,y=y)
#plt.show()
if show:
f=plt.figure()
plt.plot(cumLossgd[1:], label="Gradient Descent Loss")
f.hold(True)
plt.plot(cumLossls[1:], label="Filter Loss")
#avglossgd=np.array([cumLossgd[-1]/len(cumLossgd)]*len(cumLossgd))
#plt.plot(avglossgd,label="Average GD Loss")
#plt.plot([cumLossls[-1]/len(cumLossls)]*len(cumLossls), label="Average Filter Loss")
plt.title("CumLoss Ratio:"+str(min(cumLossgd[-1],cumLossls[-1])/max(cumLossgd[-1],cumLossls[-1])))
plt.legend()
figname="g"+str(paramlist[0])+"_nmin"+str(paramlist[1])+"_al"+str(paramlist[2])+"_thr"+str(paramlist[3])\
+ "_lr"+str(paramlist[4])+".png"
plt.savefig(figname)
#plt.show()
f.clear()
#print(i)
#print(fimtgd.count_leaves())
#print(fimtgd.count_nodes())
return [cumLossgd,cumLossls,cumLossgls,val,paramlist]
def Kiel_Test(paramlist,show,val):
fimtgd=FIMTGD(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[4])
fimtls=FIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[4])
gfimtls=gFIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[5])
cumLossgd =[0]
cumLossls =[0]
cumLossgls =[0]
if True:
data = get_Kiel_data()
c = 0
for i in range(100000):
c += 1
print(str(c)+'/'+str(100000))
input = data[i][8:10]
#target = data[1][i] + (np.random.uniform() - 0.5) * 0.2
target = data[i][10]
if i > -1:
cumLossgd.append(cumLossgd[-1] + np.fabs(target - fimtgd.eval_and_learn(np.array(input), target)))
cumLossls.append(cumLossls[-1] + np.fabs(target - fimtls.eval_and_learn(np.array(input), target)))
cumLossgls.append(cumLossgls[-1] + np.fabs(target - gfimtls.eval_and_learn(np.array(input), target)))
else:
#warm start
fimtgd.eval_and_learn(np.array(input), target)
fimtls.eval_and_learn(np.array(input), target)
gfimtls.eval_and_learn(np.array(input), target)
#plt.scatter(x=x,y=y)
#plt.show()
if show:
f=plt.figure()
plt.plot(cumLossgd[1:], label="Gradient Descent Loss")
f.hold(True)
plt.plot(cumLossls[1:], label="Filter Loss")
#avglossgd=np.array([cumLossgd[-1]/len(cumLossgd)]*len(cumLossgd))
#plt.plot(avglossgd,label="Average GD Loss")
#plt.plot([cumLossls[-1]/len(cumLossls)]*len(cumLossls), label="Average Filter Loss")
plt.title("CumLoss Ratio:"+str(min(cumLossgd[-1],cumLossls[-1])/max(cumLossgd[-1],cumLossls[-1])))
plt.legend()
figname="g"+str(paramlist[0])+"_nmin"+str(paramlist[1])+"_al"+str(paramlist[2])+"_thr"+str(paramlist[3])\
+ "_lr"+str(paramlist[4])+".png"
plt.savefig(figname)
#plt.show()
f.clear()
#print(i)
#print(fimtgd.count_leaves())
#print(fimtgd.count_nodes())
return [cumLossgd,cumLossls,cumLossgls,val,paramlist]
def abalone_test(paramlist,show,val):
#print(val)
#print(paramlist)
fimtgd=FIMTGD(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[4])
fimtls=FIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[4])
gfimtls=gFIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[5])
cumLossgd =[0]
cumLossls =[0]
cumLossgls =[0]
with open( "abalone.data", 'rt') as abalonefile:
i = 0
for row in abalonefile:
i += 1
row=row.rstrip().split(',')
target=float(row[-1])
if row[0]=="M":
numgender=1.
if row[0]=="I":
numgender=0.5
if row[0]=="F":
numgender=0.
input=[numgender]
for item in row[1:-1]:
input.append(float(item))
cumLossgd.append(cumLossgd[-1] + np.fabs(target - fimtgd.eval_and_learn(np.array(input), target)))
cumLossls.append(cumLossls[-1] + np.fabs(target - fimtls.eval_and_learn(np.array(input), target)))
cumLossgls.append(cumLossgls[-1] + np.fabs(target - gfimtls.eval_and_learn(np.array(input), target)))
if show:
f=plt.figure()
plt.plot(cumLossgd[1:], label="Gradient Descent Loss")
f.hold(True)
plt.plot(cumLossls[1:], label="Filter Loss")
#avglossgd=np.array([cumLossgd[-1]/len(cumLossgd)]*len(cumLossgd))
#plt.plot(avglossgd,label="Average GD Loss")
#plt.plot([cumLossls[-1]/len(cumLossls)]*len(cumLossls), label="Average Filter Loss")
plt.title("CumLoss Ratio:"+str(min(cumLossgd[-1],cumLossls[-1])/max(cumLossgd[-1],cumLossls[-1])))
plt.legend()
figname="g"+str(paramlist[0])+"_nmin"+str(paramlist[1])+"_al"+str(paramlist[2])+"_thr"+str(paramlist[3])\
+ "_lr"+str(paramlist[4])+".png"
plt.savefig(figname)
#plt.show()
f.clear()
#print(i)
#print(fimtgd.count_leaves())
#print(fimtgd.count_nodes())
return [cumLossgd,cumLossls,cumLossgls,val,paramlist]
def legendre_test(paramlist,show,val):
#print(val)
#print(paramlist)
fimtgd=FIMTGD(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[4])
fimtls=FIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[4])
gfimtls=gFIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=paramlist[2], threshold=paramlist[3], learn=paramlist[5])
cumLossgd =[0]
cumLossls =[0]
cumLossgls =[0]
if True:
start = 0.0
end = 1.0
i = 0
for input,target,o_target in data_provider([9,9,32,32,4],[0.05,0.05,0.05,0.05,0.05],[1000,1000,3000,2000,2000],5):
#print(i,'/',2000)
i+=1
#cumLossgd.append(cumLossgd[-1] + np.sqrt(np.fabs(o_target - fimtgd.eval_and_learn(np.array(input), target))**2))
#cumLossls.append(cumLossls[-1] + np.sqrt(np.fabs(o_target - fimtls.eval_and_learn(np.array(input), target))**2))
#cumLossgls.append(cumLossgls[-1] + np.sqrt(np.fabs(o_target - gfimtls.eval_and_learn(np.array(input), target))**2))
cumLossgd.append(cumLossgd[-1] + np.sqrt(np.fabs(o_target - fimtgd.eval_and_learn(np.array(input), target))**2))
cumLossls.append(cumLossls[-1] + np.sqrt(np.fabs(o_target - fimtls.eval_and_learn(np.array(input), target))**2))
cumLossgls.append(cumLossgls[-1] + np.sqrt(np.fabs(o_target - gfimtls.eval_and_learn(np.array(input), target))**2))
#plt.scatter(x=x,y=y)
#plt.show()
if show:
f=plt.figure()
plt.plot(cumLossgd[1:], label="Gradient Descent Loss")
f.hold(True)
plt.plot(cumLossls[1:], label="Filter Loss")
#avglossgd=np.array([cumLossgd[-1]/len(cumLossgd)]*len(cumLossgd))
#plt.plot(avglossgd,label="Average GD Loss")
#plt.plot([cumLossls[-1]/len(cumLossls)]*len(cumLossls), label="Average Filter Loss")
plt.title("CumLoss Ratio:"+str(min(cumLossgd[-1],cumLossls[-1])/max(cumLossgd[-1],cumLossls[-1])))
plt.legend()
figname="g"+str(paramlist[0])+"_nmin"+str(paramlist[1])+"_al"+str(paramlist[2])+"_thr"+str(paramlist[3])\
+ "_lr"+str(paramlist[4])+".png"
plt.savefig(figname)
#plt.show()
f.clear()
#print(i)
#print(fimtgd.count_leaves())
#print(fimtgd.count_nodes())
return [cumLossgd,cumLossls,cumLossgls,val,paramlist]
def test2d(paramlist,show,val):
#print(val)
#print(paramlist)
fimtgd=FIMTGD(gamma=paramlist[0], n_min = paramlist[1], alpha=[2], threshold=paramlist[3], learn=paramlist[4])
fimtls=FIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=[2], threshold=paramlist[3], learn=paramlist[4])
gfimtls=gFIMTLS(gamma=paramlist[0], n_min = paramlist[1], alpha=[2], threshold=paramlist[3], learn=paramlist[5])
cumLossgd =[0]
cumLossls =[0]
cumLossgls =[0]
if True:
start = 0.0
end = 1.0
X = list()
Y = list()
x, y, z = axes3d.get_test_data(0.1)
num_d = len(x)**2
for i in range(len(x)):
for j in range(len(y)):
input = [x[i,j],y[i,j]]
target = z[i,j]
X.append(input)
Y.append(target)
data = [X,Y]
data = np.array(data)
data = data.transpose()
np.random.shuffle(data)
data = data.transpose()
for i in range(num_d):
input = data[0][i]
target = data[1][i] + (np.random.uniform() - 0.5) * 0.2
o_target = data[1][i]
if num_d/2 < i:
target += 1.0
o_target += 1.0
cumLossgd.append(cumLossgd[-1] + np.fabs(o_target - fimtgd.eval_and_learn(np.array(input), target)))
cumLossls.append(cumLossls[-1] + np.fabs(o_target - fimtls.eval_and_learn(np.array(input), target)))
cumLossgls.append(cumLossgls[-1] + np.fabs(o_target - gfimtls.eval_and_learn(np.array(input), target)))
#plt.scatter(x=x,y=y)
#plt.show()
if show:
f=plt.figure()
plt.plot(cumLossgd[1:], label="Gradient Descent Loss")
f.hold(True)
plt.plot(cumLossls[1:], label="Filter Loss")
#avglossgd=np.array([cumLossgd[-1]/len(cumLossgd)]*len(cumLossgd))
#plt.plot(avglossgd,label="Average GD Loss")
#plt.plot([cumLossls[-1]/len(cumLossls)]*len(cumLossls), label="Average Filter Loss")
plt.title("CumLoss Ratio:"+str(min(cumLossgd[-1],cumLossls[-1])/max(cumLossgd[-1],cumLossls[-1])))
plt.legend()
figname="g"+str(paramlist[0])+"_nmin"+str(paramlist[1])+"_al"+str(paramlist[2])+"_thr"+str(paramlist[3])\
+ "_lr"+str(paramlist[4])+".png"
plt.savefig(figname)
#plt.show()
f.clear()
#print(i)
#print(fimtgd.count_leaves())
#print(fimtgd.count_nodes())
return [cumLossgd,cumLossls,cumLossgls,val,paramlist]