auto with optim instead of grid
This commit is contained in:
parent
2aabfdc4c0
commit
9020d7ff31
|
|
@ -271,7 +271,7 @@ class KDEyMLauto2(KDEyML):
|
|||
self.reduction = reduction
|
||||
self.max_reduced = max_reduced
|
||||
self.random_state = random_state
|
||||
assert target == 'likelihood' or target in qp.error.QUANTIFICATION_ERROR_NAMES, 'unknown target for auto'
|
||||
assert target in ['likelihood', 'likelihood+'] or target in qp.error.QUANTIFICATION_ERROR_NAMES, 'unknown target for auto'
|
||||
self.target = target
|
||||
|
||||
def aggregation_fit(self, classif_predictions: LabelledCollection, data: LabelledCollection):
|
||||
|
|
@ -293,34 +293,60 @@ class KDEyMLauto2(KDEyML):
|
|||
if len(train) > tr_length:
|
||||
train = train.sampling(tr_length)
|
||||
|
||||
best_band = None
|
||||
best_loss_val = None
|
||||
init_prev = np.full(fill_value=1 / n_classes, shape=(n_classes,))
|
||||
for bandwidth in np.logspace(-4, np.log10(0.2), 20):
|
||||
mix_densities = self.get_mixture_components(*train.Xy, train.classes_, bandwidth)
|
||||
repeats = 25
|
||||
prot = UPP(val, sample_size=self.reduction, repeats=repeats, random_state=self.random_state)
|
||||
|
||||
repeats = 25
|
||||
loss_accum = 0
|
||||
prot = UPP(val, sample_size=self.reduction, repeats=repeats, random_state=self.random_state)
|
||||
for (sample, prev) in tqdm(prot(), total=repeats):
|
||||
test_densities = [self.pdf(kde_i, sample) for kde_i in mix_densities]
|
||||
if self.target == 'likelihood+':
|
||||
def neg_loglikelihood_band_(bandwidth):
|
||||
mix_densities = self.get_mixture_components(*train.Xy, train.classes_, bandwidth)
|
||||
loss_accum = 0
|
||||
|
||||
def neg_loglikelihood_prev_(prev):
|
||||
test_mixture_likelihood = sum(prev_i * dens_i for prev_i, dens_i in zip(prev, test_densities))
|
||||
test_loglikelihood = np.log(test_mixture_likelihood + epsilon)
|
||||
return -np.sum(test_loglikelihood)
|
||||
for (sample, prev) in tqdm(prot(), total=repeats):
|
||||
test_densities = [self.pdf(kde_i, sample) for kde_i in mix_densities]
|
||||
|
||||
if self.target == 'likelihood':
|
||||
loss_fn = neg_loglikelihood_prev_
|
||||
else:
|
||||
loss_fn = lambda prev_hat: qp.error.from_name(self.target)(prev, prev_hat)
|
||||
def neg_loglikelihood_prev_(prev):
|
||||
test_mixture_likelihood = sum(prev_i * dens_i for prev_i, dens_i in zip(prev, test_densities))
|
||||
test_loglikelihood = np.log(test_mixture_likelihood + epsilon)
|
||||
return -np.sum(test_loglikelihood)
|
||||
|
||||
pred_prev, loss_val = optim_minimize(loss_fn, init_prev, return_loss=True)
|
||||
loss_accum += loss_val
|
||||
pred_prev, loss_val = optim_minimize(neg_loglikelihood_prev_, init_prev, return_loss=True)
|
||||
loss_accum += loss_val
|
||||
|
||||
if best_loss_val is None or loss_accum < best_loss_val:
|
||||
best_loss_val = loss_accum
|
||||
best_band = bandwidth
|
||||
return loss_accum
|
||||
|
||||
bounds = [tuple(0, 1)]
|
||||
init_bandwidth = 0.1
|
||||
r = optimize.minimize(neg_loglikelihood_band_, x0=[init_bandwidth], method='SLSQP', bounds=bounds)
|
||||
best_band = r.x[0]
|
||||
|
||||
else:
|
||||
best_band = None
|
||||
best_loss_val = None
|
||||
init_prev = np.full(fill_value=1 / n_classes, shape=(n_classes,))
|
||||
for bandwidth in np.logspace(-4, np.log10(0.2), 20):
|
||||
mix_densities = self.get_mixture_components(*train.Xy, train.classes_, bandwidth)
|
||||
|
||||
loss_accum = 0
|
||||
for (sample, prev) in tqdm(prot(), total=repeats):
|
||||
test_densities = [self.pdf(kde_i, sample) for kde_i in mix_densities]
|
||||
|
||||
def neg_loglikelihood_prev_(prev):
|
||||
test_mixture_likelihood = sum(prev_i * dens_i for prev_i, dens_i in zip(prev, test_densities))
|
||||
test_loglikelihood = np.log(test_mixture_likelihood + epsilon)
|
||||
return -np.sum(test_loglikelihood)
|
||||
|
||||
if self.target == 'likelihood':
|
||||
loss_fn = neg_loglikelihood_prev_
|
||||
else:
|
||||
loss_fn = lambda prev_hat: qp.error.from_name(self.target)(prev, prev_hat)
|
||||
|
||||
pred_prev, loss_val = optim_minimize(loss_fn, init_prev, return_loss=True)
|
||||
loss_accum += loss_val
|
||||
|
||||
if best_loss_val is None or loss_accum < best_loss_val:
|
||||
best_loss_val = loss_accum
|
||||
best_band = bandwidth
|
||||
|
||||
print(f'found bandwidth={best_band:.4f} (loss_val={best_loss_val:.5f})')
|
||||
self.bandwidth_ = best_band
|
||||
|
|
|
|||
|
|
@ -38,6 +38,7 @@ METHODS = [
|
|||
('KDEy-ML-scott', KDEyML(newLR(), bandwidth='scott'), wrap_hyper(logreg_grid)),
|
||||
('KDEy-ML-silver', KDEyML(newLR(), bandwidth='silverman'), wrap_hyper(logreg_grid)),
|
||||
('KDEy-ML-autoLike', KDEyMLauto2(newLR(), bandwidth='auto', target='likelihood'), wrap_hyper(logreg_grid)),
|
||||
('KDEy-ML-autoLike+', KDEyMLauto2(newLR(), bandwidth='auto', target='likelihood+'), wrap_hyper(logreg_grid)),
|
||||
('KDEy-ML-autoAE', KDEyMLauto2(newLR(), bandwidth='auto', target='mae'), wrap_hyper(logreg_grid)),
|
||||
('KDEy-ML-autoRAE', KDEyMLauto2(newLR(), bandwidth='auto', target='mrae'), wrap_hyper(logreg_grid)),
|
||||
]
|
||||
|
|
|
|||
|
|
@ -14,6 +14,8 @@ from quapy.model_selection import GridSearchQ
|
|||
from quapy.protocol import UPP
|
||||
from pathlib import Path
|
||||
from quapy import functional as F
|
||||
import matplotlib.pyplot as plt
|
||||
|
||||
|
||||
SEED = 1
|
||||
|
||||
|
|
@ -24,84 +26,96 @@ def newLR():
|
|||
SAMPLE_SIZE=150
|
||||
qp.environ['SAMPLE_SIZE'] = SAMPLE_SIZE
|
||||
|
||||
show_ae = True
|
||||
show_rae = True
|
||||
show_mse = False
|
||||
show_kld = True
|
||||
|
||||
epsilon = 1e-10
|
||||
# n_bags_test = 2
|
||||
# DATASETS = [qp.datasets.UCI_MULTICLASS_DATASETS[21]]
|
||||
DATASETS = qp.datasets.UCI_MULTICLASS_DATASETS
|
||||
for i, dataset in enumerate(DATASETS):
|
||||
data = qp.datasets.fetch_UCIMulticlassDataset(dataset)
|
||||
n_classes = data.n_classes
|
||||
print(f'{i=}')
|
||||
print(f'{dataset=}')
|
||||
print(f'{n_classes=}')
|
||||
print(len(data.training))
|
||||
print(len(data.test))
|
||||
|
||||
train, test = data.train_test
|
||||
train_prev = train.prevalence()
|
||||
test_prev = test.prevalence()
|
||||
def generate_data():
|
||||
data = qp.datasets.fetch_UCIMulticlassDataset(dataset)
|
||||
n_classes = data.n_classes
|
||||
print(f'{i=}')
|
||||
print(f'{dataset=}')
|
||||
print(f'{n_classes=}')
|
||||
print(len(data.training))
|
||||
print(len(data.test))
|
||||
|
||||
print(f'train-prev = {F.strprev(train_prev)}')
|
||||
print(f'test-prev = {F.strprev(test_prev)}')
|
||||
train, test = data.train_test
|
||||
train_prev = train.prevalence()
|
||||
test_prev = test.prevalence()
|
||||
|
||||
# protocol = UPP(test, repeats=n_bags_test)
|
||||
#
|
||||
# for sample, prev in protocol():
|
||||
# print(f'sample-prev = {F.strprev(prev)}')
|
||||
print(f'train-prev = {F.strprev(train_prev)}')
|
||||
print(f'test-prev = {F.strprev(test_prev)}')
|
||||
|
||||
# prev = np.asarray([0.2, 0.3, 0.5])
|
||||
# prev = np.asarray([0.33, 0.33, 0.34])
|
||||
# prev = train_prev
|
||||
repeats = 10
|
||||
prot = UPP(test, sample_size=SAMPLE_SIZE, repeats=repeats)
|
||||
kde = KDEyMLauto(newLR())
|
||||
kde.fit(train)
|
||||
AE_error, RAE_error, MSE_error, KLD_error, LIKE_value = [], [], [], [], []
|
||||
tr_posteriors, tr_y = kde.classif_predictions.Xy
|
||||
for it, (sample, prev) in tqdm(enumerate(prot()), total=repeats):
|
||||
te_posteriors = kde.classifier.predict_proba(sample)
|
||||
classes = train.classes_
|
||||
|
||||
# sample = test.sampling(SAMPLE_SIZE, *prev, random_state=1)
|
||||
# print(f'sample-prev = {F.strprev(prev)}')
|
||||
xaxis = []
|
||||
ae_error = []
|
||||
rae_error = []
|
||||
mse_error = []
|
||||
kld_error = []
|
||||
likelihood_value = []
|
||||
|
||||
repeats = 10
|
||||
prot = UPP(test, sample_size=SAMPLE_SIZE, repeats=repeats)
|
||||
kde = KDEyMLauto(newLR())
|
||||
kde.fit(train)
|
||||
tr_posteriors, tr_y = kde.classif_predictions.Xy
|
||||
for it, (sample, prev) in tqdm(enumerate(prot()), total=repeats):
|
||||
te_posteriors = kde.classifier.predict_proba(sample)
|
||||
classes = train.classes_
|
||||
# for bandwidth in np.linspace(0.01, 0.2, 50):
|
||||
for bandwidth in np.logspace(-5, 0.5, 50):
|
||||
mix_densities = kde.get_mixture_components(tr_posteriors, tr_y, classes, bandwidth)
|
||||
test_densities = [kde.pdf(kde_i, te_posteriors) for kde_i in mix_densities]
|
||||
|
||||
xaxis = []
|
||||
ae_error = []
|
||||
rae_error = []
|
||||
mse_error = []
|
||||
kld_error = []
|
||||
likelihood_val = []
|
||||
def neg_loglikelihood_prev(prev):
|
||||
test_mixture_likelihood = sum(prev_i * dens_i for prev_i, dens_i in zip(prev, test_densities))
|
||||
test_loglikelihood = np.log(test_mixture_likelihood + epsilon)
|
||||
return -np.sum(test_loglikelihood)
|
||||
|
||||
# for bandwidth in np.linspace(0.01, 0.2, 50):
|
||||
for bandwidth in np.logspace(-3, 0.5, 50):
|
||||
mix_densities = kde.get_mixture_components(tr_posteriors, tr_y, classes, bandwidth)
|
||||
test_densities = [kde.pdf(kde_i, te_posteriors) for kde_i in mix_densities]
|
||||
init_prev = np.full(fill_value=1 / n_classes, shape=(n_classes,))
|
||||
pred_prev, likelihood = optim_minimize(neg_loglikelihood_prev, init_prev, return_loss=True)
|
||||
|
||||
def neg_loglikelihood_prev(prev):
|
||||
test_mixture_likelihood = sum(prev_i * dens_i for prev_i, dens_i in zip(prev, test_densities))
|
||||
test_loglikelihood = np.log(test_mixture_likelihood + epsilon)
|
||||
return -np.sum(test_loglikelihood)
|
||||
xaxis.append(bandwidth)
|
||||
ae_error.append(qp.error.ae(prev, pred_prev))
|
||||
rae_error.append(qp.error.rae(prev, pred_prev))
|
||||
mse_error.append(qp.error.mse(prev, pred_prev))
|
||||
kld_error.append(qp.error.kld(prev, pred_prev))
|
||||
likelihood_value.append(likelihood)
|
||||
|
||||
init_prev = np.full(fill_value=1 / n_classes, shape=(n_classes,))
|
||||
pred_prev, likelihood = optim_minimize(neg_loglikelihood_prev, init_prev, return_loss=True)
|
||||
AE_error.append(ae_error)
|
||||
RAE_error.append(rae_error)
|
||||
MSE_error.append(mse_error)
|
||||
KLD_error.append(kld_error)
|
||||
LIKE_value.append(likelihood_value)
|
||||
|
||||
xaxis.append(bandwidth)
|
||||
ae_error.append(qp.error.ae(prev, pred_prev))
|
||||
rae_error.append(qp.error.rae(prev, pred_prev))
|
||||
mse_error.append(qp.error.mse(prev, pred_prev))
|
||||
kld_error.append(qp.error.kld(prev, pred_prev))
|
||||
likelihood_val.append(likelihood)
|
||||
return xaxis, AE_error, RAE_error, MSE_error, KLD_error, LIKE_value
|
||||
|
||||
import matplotlib.pyplot as plt
|
||||
xaxis, AE_error, RAE_error, MSE_error, KLD_error, LIKE_value = qp.util.pickled_resource(
|
||||
f'./plots/likelihood/pickles/{dataset}.pkl', generate_data)
|
||||
|
||||
for row in range(len(AE_error)):
|
||||
|
||||
# Crear la figura
|
||||
# ----------------------------------------------------------------------------------------------------
|
||||
fig, ax1 = plt.subplots(figsize=(8, 6))
|
||||
|
||||
# Pintar las series ae_error, rae_error, y kld_error en el primer eje Y
|
||||
ax1.plot(xaxis, ae_error, label='AE Error', marker='o', color='b')
|
||||
# ax1.plot(xaxis, rae_error, label='RAE Error', marker='s', color='g')
|
||||
# ax1.plot(xaxis, kld_error, label='KLD Error', marker='^', color='r')
|
||||
ax1.plot(xaxis, mse_error, label='MSE Error', marker='^', color='c')
|
||||
if show_ae:
|
||||
ax1.plot(xaxis, AE_error[row], label='AE', marker='o', color='b')
|
||||
if show_rae:
|
||||
ax1.plot(xaxis, RAE_error[row], label='RAE', marker='s', color='g')
|
||||
if show_kld:
|
||||
ax1.plot(xaxis, KLD_error[row], label='KLD', marker='^', color='r')
|
||||
if show_mse:
|
||||
ax1.plot(xaxis, MSE_error[row], label='MSE', marker='^', color='c')
|
||||
ax1.set_xscale('log')
|
||||
|
||||
# Configurar etiquetas para el primer eje Y
|
||||
|
|
@ -114,7 +128,7 @@ for i, dataset in enumerate(DATASETS):
|
|||
ax2 = ax1.twinx()
|
||||
|
||||
# Pintar likelihood_val en el segundo eje Y
|
||||
ax2.plot(xaxis, likelihood_val, label='(neg)Likelihood', marker='x', color='purple')
|
||||
ax2.plot(xaxis, LIKE_value[row], label='(neg)Likelihood', marker='x', color='purple')
|
||||
|
||||
# Configurar etiquetas para el segundo eje Y
|
||||
ax2.set_ylabel('Likelihood Value')
|
||||
|
|
@ -124,9 +138,50 @@ for i, dataset in enumerate(DATASETS):
|
|||
plt.title('Error Metrics vs Bandwidth')
|
||||
# plt.show()
|
||||
os.makedirs('./plots/likelihood/', exist_ok=True)
|
||||
plt.savefig(f'./plots/likelihood/{dataset}-fig{it}.png')
|
||||
plt.savefig(f'./plots/likelihood/{dataset}-fig{row}.png')
|
||||
plt.close()
|
||||
|
||||
# Crear la figura con las medias
|
||||
# ----------------------------------------------------------------------------------------------------
|
||||
fig, ax1 = plt.subplots(figsize=(8, 6))
|
||||
|
||||
def add_plot(ax, vals_error, name, color, marker, show):
|
||||
if not show:
|
||||
return
|
||||
vals_error = np.asarray(vals_error)
|
||||
vals_ave = np.mean(vals_error, axis=0)
|
||||
vals_std = np.std(vals_error, axis=0)
|
||||
ax.plot(xaxis, vals_ave, label=name, marker=marker, color=color)
|
||||
ax.fill_between(xaxis, vals_ave - vals_std, vals_ave + vals_std, color=color, alpha=0.2)
|
||||
|
||||
add_plot(ax1, AE_error, 'AE', color='b', marker='o', show=show_ae)
|
||||
add_plot(ax1, RAE_error, 'RAE', color='g', marker='s', show=show_rae)
|
||||
add_plot(ax1, KLD_error, 'KLD', color='r', marker='^', show=show_kld)
|
||||
add_plot(ax1, MSE_error, 'MSE', color='c', marker='^', show=show_mse)
|
||||
ax1.set_xscale('log')
|
||||
|
||||
# Configurar etiquetas para el primer eje Y
|
||||
ax1.set_xlabel('Bandwidth')
|
||||
ax1.set_ylabel('Error Value')
|
||||
ax1.grid(True)
|
||||
ax1.legend(loc='upper left')
|
||||
|
||||
# Crear un segundo eje Y que comparte el eje X
|
||||
ax2 = ax1.twinx()
|
||||
|
||||
# Pintar likelihood_val en el segundo eje Y
|
||||
add_plot(ax2, LIKE_value, '(neg)Likelihood', color='purple', marker='x', show=True)
|
||||
|
||||
# Configurar etiquetas para el segundo eje Y
|
||||
ax2.set_ylabel('Likelihood Value')
|
||||
ax2.legend(loc='upper right')
|
||||
|
||||
# Mostrar el gráfico
|
||||
plt.title('Error Metrics vs Bandwidth')
|
||||
# plt.show()
|
||||
os.makedirs('./plots/likelihood/', exist_ok=True)
|
||||
plt.savefig(f'./plots/likelihood/{dataset}-figAve.png')
|
||||
plt.close()
|
||||
|
||||
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue