def tree_loss_method(args,
model_noisy,
y_train_noisy,
noisy_indices,
n_check,
n_checkpoint,
clf,
X_train,
y_train,
X_test,
y_test,
acc_noisy,
auc_noisy,
logger=None):
"""
Orders training instances by largest loss.
"""
y_train_proba = model_noisy.predict_proba(X_train)[:, 1]
y_train_noisy_loss = util.instance_log_loss(
y_train_noisy, y_train_proba) # negative log-likelihood
train_indices = np.argsort(
y_train_noisy_loss)[::-1] # descending, largest log loss first
result = fix_noisy_instances(train_indices,
noisy_indices,
n_check,
n_checkpoint,
clf,
X_train,
y_train,
X_test,
y_test,
acc_noisy,
auc_noisy,
logger=logger)
return result
def teknn_method(args, model_noisy, y_train_noisy,
noisy_indices, n_check, n_checkpoint,
clf, X_train, y_train, X_test, y_test,
acc_noisy, auc_noisy, logger=None,
frac_progress_update=0.1):
"""
Count impact by the number of times a training sample shows up in
one another's neighborhood, this can be weighted by 1 / distance.
"""
# train surrogate model
params = {'C': args.C, 'n_neighbors': args.n_neighbors, 'tree_kernel': args.tree_kernel}
train_label = y_train_noisy if 'og' in args.method else model_noisy.predict(X_train)
surrogate = trex.train_surrogate(model=model_noisy,
surrogate=args.method.split('_')[0],
X_train=X_train,
y_train=train_label,
val_frac=args.tune_frac,
metric=args.metric,
seed=args.rs,
params=params,
logger=logger)
# sort by instance log loss using the surrogate model
if 'loss' in args.method:
# display progress
if logger:
logger.info('\ncomputing KNN loss...')
y_train_proba = surrogate.predict_proba(X_train)[:, 1]
y_train_noisy_loss = util.instance_log_loss(y_train_noisy, y_train_proba) # negative log-likelihood
train_indices = np.argsort(y_train_noisy_loss)[::-1] # descending, largest log loss first
# sort by absolute value of instance weights
else:
# sort instances based on largest influence toward the predicted training labels
attributions = surrogate.compute_attributions(X_train, logger=logger)
attributions_sum = np.sum(attributions, axis=0)
train_indices = np.argsort(attributions_sum)[::-1]
# fix noisy instances
result = fix_noisy_instances(train_indices, noisy_indices, n_check, n_checkpoint,
clf, X_train, y_train, X_test, y_test,
acc_noisy, auc_noisy, logger=logger)
return result
def trex_method(args, model_noisy, y_train_noisy,
noisy_indices, n_check, n_checkpoint,
clf, X_train, y_train, X_test, y_test,
acc_noisy, auc_noisy, logger=None):
"""
Order by largest absolute values of the instance coefficients
from the KLR or SVM surrogate model.
"""
# train surrogate model
params = {'C': args.C, 'n_neighbors': args.n_neighbors, 'tree_kernel': args.tree_kernel}
train_label = y_train_noisy if 'og' in args.method else model_noisy.predict(X_train)
surrogate = trex.train_surrogate(model=model_noisy,
surrogate=args.method.split('_')[0],
X_train=X_train,
y_train=train_label,
val_frac=args.tune_frac,
metric=args.metric,
seed=args.rs,
params=params,
logger=logger)
# sort by instance log loss using the surrogate model
if 'loss' in args.method:
y_train_proba = surrogate.predict_proba(X_train)[:, 1]
y_train_noisy_loss = util.instance_log_loss(y_train_noisy, y_train_proba) # negative log-likelihood
train_indices = np.argsort(y_train_noisy_loss)[::-1] # descending, largest log loss first
# sort by absolute value of instance weights
else:
train_indices = np.argsort(np.abs(surrogate.get_alpha()))[::-1]
# fix noisy instances
result = fix_noisy_instances(train_indices, noisy_indices, n_check, n_checkpoint,
clf, X_train, y_train, X_test, y_test,
acc_noisy, auc_noisy, logger=logger)
return result
def trex_method(args,
model_noisy,
y_train_noisy,
noisy_indices,
n_check,
n_checkpoint,
clf,
X_train,
y_train,
X_test,
y_test,
acc_noisy,
auc_noisy,
logger=None,
out_dir=None):
"""
Order by largest absolute values of the instance coefficients
from the KLR or SVM surrogate model.
"""
# train surrogate model
params = util.get_selected_params(dataset=args.dataset,
model=args.model,
surrogate=args.method)
train_label = y_train_noisy if 'og' in args.method else model_noisy.predict(
X_train)
surrogate = trex.train_surrogate(model=model_noisy,
surrogate='klr',
X_train=X_train,
y_train=train_label,
val_frac=args.tune_frac,
metric=args.metric,
seed=args.rs,
params=params,
logger=None)
# sort by instance log loss using the surrogate model
if 'loss' in args.method:
y_train_proba = surrogate.predict_proba(X_train)[:, 1]
y_train_noisy_loss = util.instance_log_loss(
y_train_noisy, y_train_proba) # negative log-likelihood
train_indices = np.argsort(
y_train_noisy_loss)[::-1] # descending, largest log loss first
# sort by sim. or alpha
else:
logger.info('\nsorting train instances...')
# sort by similarity
if 'sim' in args.method:
start = time.time()
# prioritize largest absolute similarity density
if 'abs' in args.method:
similarity_density = np.zeros(0, )
n_chunk = int(X_train.shape[0] * 0.1)
for i in range(0, X_train.shape[0], n_chunk):
X_sub_sim = surrogate.similarity(X_train[i:i + n_chunk])
similarity_density_sub = np.sum(X_sub_sim, axis=1)
similarity_density = np.concatenate(
[similarity_density, similarity_density_sub])
elapsed = time.time() - start
logger.info('{:.1f}%...{:.3f}s'.format(
i / X_train.shape[0] * 100, elapsed))
similarity_density = np.abs(similarity_density)
train_indices = np.argsort(similarity_density)[::-1]
# sort train instances prioritizing largest negative similarity density
else:
similarity_density = np.zeros(0, )
n_chunk = int(X_train.shape[0] * 0.1)
for i in range(0, X_train.shape[0], n_chunk):
X_sub_sim = surrogate.similarity(X_train[i:i + n_chunk])
y_sub_mask = np.ones(X_sub_sim.shape)
for j in range(y_sub_mask.shape[0]):
y_sub_mask[j][np.where(
y_train_noisy[j + i] != y_train_noisy)] = -1
X_sub_sim = X_sub_sim * y_sub_mask
similarity_density_sub = np.sum(X_sub_sim, axis=1)
similarity_density = np.concatenate(
[similarity_density, similarity_density_sub])
elapsed = time.time() - start
logger.info('{:.1f}%...{:.3f}s'.format(
i / X_train.shape[0] * 100, elapsed))
train_indices = np.argsort(similarity_density)
# plot |alpha| vs. similarity density
if out_dir is not None:
alpha = surrogate.get_alpha()
alpha = np.abs(alpha)
non_noisy_indices = np.setdiff1d(np.arange(y_train.shape[0]),
noisy_indices)
fig, ax = plt.subplots()
ax.scatter(alpha[non_noisy_indices],
similarity_density[non_noisy_indices],
alpha=0.1,
label='non-noisy',
color='green')
ax.scatter(alpha[noisy_indices],
similarity_density[noisy_indices],
alpha=0.1,
label='noisy',
color='red')
ax.set_xlabel('alpha')
ax.set_ylabel('similarity_density')
ax.legend()
plt.savefig(os.path.join(out_dir, 'alpha_sim.png'))
elif 'alpha' in args.method:
alpha = surrogate.get_alpha()
magnitude = np.abs(alpha)
train_indices = np.argsort(magnitude)[::-1]
else:
raise ValueError('unknown method {}'.format(args.method))
# fix noisy instances
result = fix_noisy_instances(train_indices,
noisy_indices,
n_check,
n_checkpoint,
clf,
X_train,
y_train,
X_test,
y_test,
acc_noisy,
auc_noisy,
logger=logger)
return result