def remove_exs(dataset, hyps, err_hyp, alpha, split_name, one_v_all):
# only keep examples that we can predict with the best hypothesis
if one_v_all:
if np.unique(dataset['Y'].shape[0]) == 2:
# binary
optimal_index = np.argmin(err_hyp[0])
_, pred_class = teach.user_model_binary(hyps[optimal_index],
dataset['X'], dataset['Y'],
alpha)
inds = np.where(dataset['Y'] == pred_class)[0]
else:
# multi class
correctly_predicted = np.zeros(dataset['Y'].shape[0])
for cc in range(len(err_hyp)):
optimal_index = np.argmin(err_hyp[cc])
Y_bin = np.zeros(dataset['Y'].shape[0]).astype(np.int)
Y_bin[np.where(dataset['Y'] == cc)[0]] = 1
_, pred_class = teach.user_model_binary(
hyps[optimal_index], dataset['X'], Y_bin, alpha)
correctly_predicted[np.where(Y_bin == pred_class)[0]] += 1
inds = np.where(correctly_predicted == len(err_hyp))[0]
else:
optimal_index = np.argmin(err_hyp)
_, pred_class = teach.user_model(hyps[optimal_index], dataset['X'],
dataset['Y'], alpha)
inds = np.where(dataset['Y'] == pred_class)[0]
print(dataset['X'].shape[0] - inds.shape[0], split_name,
'examples removed')
# remove the examples
dataset['X'] = dataset['X'][inds, :]
dataset['Y'] = dataset['Y'][inds]
dataset['im_files'] = dataset['im_files'][inds]
dataset['explain_files'] = dataset['explain_files'][inds]
dataset['explain_interp'] = dataset['explain_interp'][inds]
cls_un, cls_cnt = np.unique(dataset['Y'], return_counts=True)
if 'X_density' in list(dataset.keys()):
dataset['X_density'] = dataset['X_density'][inds]
dataset['X_distance'] = dataset['X_distance'][inds, :]
dataset['X_distance'] = dataset['X_distance'][:, inds]
print('\n', split_name)
for cc in range(len(cls_cnt)):
print(cls_un[cc], dataset['class_names'][cls_un[cc]].ljust(30), '\t',
cls_cnt[cc])
return dataset
def compute_hyps_error(hyps, X, Y, alpha, one_v_all=False):
# compute err(h, h*) - vector of length H
err_hyp = np.zeros(len(hyps))
for hh in range(len(hyps)):
if one_v_all:
_, pred_class = teach.user_model_binary(hyps[hh], X, Y, alpha)
else:
_, pred_class = teach.user_model(hyps[hh], X, Y, alpha)
err_hyp[hh] = (Y != pred_class).sum() / float(Y.shape[0])
return err_hyp
def compute_likelihood(hyps, X, Y, alpha, one_v_all=False):
# compute P(y|h,x) - size HxN
# is set to one where h(x) = y i.e. correct guess
likelihood = np.ones((len(hyps), X.shape[0]))
likelihood_opp = np.ones((len(hyps), X.shape[0]))
for hh in range(len(hyps)):
if one_v_all:
# assumes that hyps[hh] is a D dim vector
prob_agree, pred_class = teach.user_model_binary(hyps[hh], X, Y, alpha)
else:
# assumes that hyps[hh] is a CxD dim maxtrix
prob_agree, pred_class = teach.user_model(hyps[hh], X, Y, alpha)
inds = np.where(pred_class != Y)[0]
likelihood[hh, inds] = prob_agree[inds]
return likelihood