(all_cat_lbls['train'], all_attr_lbls['train']))
bin_labels_test = np.hstack((all_cat_lbls['test'], all_attr_lbls['test']))
print 'num train', bin_labels_train.shape[0]
print 'num test', bin_labels_test.shape[0]
res = {}
ela_types = {'rand', 'pop', 'backoff', 'dist'}
for et in ela_types:
print '***** %s *****' % et
res[et] = {}
dtree = {}
for numq in [10, 20, 30, 40, 50, 80, 100]:
sgraph = cooccurrence.SGraph(train=bin_labels_train,
dtree=dtree,
ela_type=et,
ela_limit_type='numq',
ela_limit=numq)
res[et][numq] = []
for ind, row in enumerate(bin_labels_test[:100]):
item = sgraph.test(row, known_inds=range(num_cat))
res[et][numq].append(item[2])
print 'numq = %d: rec %d = %.2f' % (numq, ind, item[2])
dtree = sgraph.dtree
# Need to go back through and look for rare attributes
# TODO: save ELA labeled instances for each major marker
# for M 'rarest' attributes
bin_labels_train = bin_labels[:7000][:]
bin_labels_test = bin_labels[7000:8000][:]
# res = {}
# ela_types = ['rand', 'pop', 'dist'] # 'backoff',
res = joblib.load('data/sun_attr_rec_benchmark_mle_et_threshold.jbl')
ela_types = ['backoff']
for et in ela_types:
print '***** %s *****' % et
res[et] = {}
for thresh in np.arange(0.005, 0.055, 0.005):
dtree = {}
print 'threshold : ' + str(thresh)
sgraph = cooccurrence.SGraph(train=bin_labels_train,
dtree=dtree,
ela_type=et,
ela_limit_type='threshold',
ela_limit=thresh)
res[et][thresh] = {}
res[et][thresh]['rec'] = []
res[et][thresh]['numq'] = []
for ind, row in enumerate(bin_labels_test[:100][:]):
item = sgraph.test(row)
res[et][thresh]['rec'].append(item[2])
res[et][thresh]['numq'].append(len(item[3]))
print 'thresh = %.3f: rec %d = %.2f, numq = %d' % (
thresh, ind, item[2], len(item[3]))
dtree = sgraph.dtree