def eval_verification(descr, split):
print('>> Evaluating %s task' % green('verification'))
start = time.time()
pos = pd.read_csv(
os.path.join(tskdir,
'verif_pos_split-' + split['name'] + '.csv')).as_matrix()
neg_intra = pd.read_csv(
os.path.join(tskdir, 'verif_neg_intra_split-' + split['name'] +
'.csv')).as_matrix()
neg_inter = pd.read_csv(
os.path.join(tskdir, 'verif_neg_inter_split-' + split['name'] +
'.csv')).as_matrix()
d_pos = get_verif_dists(descr, pos, 1)
d_neg_intra = get_verif_dists(descr, neg_intra, 2)
d_neg_inter = get_verif_dists(descr, neg_inter, 3)
results = defaultdict(lambda: defaultdict(lambda: defaultdict(dict)))
for t in tp:
l = np.vstack((np.zeros_like(d_pos[t]), np.ones_like(d_pos[t])))
d_intra = np.vstack((d_neg_intra[t], d_pos[t]))
d_inter = np.vstack((d_neg_inter[t], d_pos[t]))
# get results for the balanced protocol: 1M Positives - 1M Negatives
fpr, tpr, auc = metrics.roc(-d_intra, l)
results[t]['intra']['balanced']['fpr'] = fpr
results[t]['intra']['balanced']['tpr'] = tpr
results[t]['intra']['balanced']['auc'] = auc
fpr, tpr, auc = metrics.roc(-d_inter, l)
results[t]['inter']['balanced']['fpr'] = fpr
results[t]['inter']['balanced']['tpr'] = tpr
results[t]['inter']['balanced']['auc'] = auc
# get results for the imbalanced protocol: 0.2M Positives - 1M Negatives
N_imb = d_pos[t].shape[0] + int(d_pos[t].shape[0] * 0.2) # 1M + 0.2*1M
pr, rc, ap = metrics.pr(-d_intra[0:N_imb], l[0:N_imb])
results[t]['intra']['imbalanced']['pr'] = pr
results[t]['intra']['imbalanced']['rc'] = rc
results[t]['intra']['imbalanced']['ap'] = ap
pr, rc, ap = metrics.pr(-d_inter[0:N_imb], l[0:N_imb])
results[t]['inter']['imbalanced']['pr'] = pr
results[t]['inter']['imbalanced']['rc'] = rc
results[t]['inter']['imbalanced']['ap'] = ap
end = time.time()
print(">> %s task finished in %.0f secs " %
(green('Verification'), end - start))
return results
def eval_matching(descr, split):
print('>> Evaluating %s task' % green('matching'))
start = time.time()
results = defaultdict(lambda: defaultdict(lambda: defaultdict(dict)))
pbar = tqdm(split['test'])
for seq in pbar:
d_ref = getattr(descr[seq], 'ref')
gt_l = np.arange(d_ref.shape[0])
for t in tp:
for i in range(1, 6):
d = getattr(descr[seq], t + str(i))
D = dist_matrix(d_ref, d, descr['distance'])
idx = np.argmin(D, axis=1)
m_l = np.equal(idx, gt_l)
results[seq][t][i]['sr'] = np.count_nonzero(m_l) / float(
m_l.shape[0])
m_d = D[gt_l, idx]
pr, rc, ap = metrics.pr(-m_d, m_l, numpos=m_l.shape[0])
results[seq][t][i]['ap'] = ap
results[seq][t][i]['pr'] = pr
results[seq][t][i]['rc'] = rc
# print(t,i,ap,results[seq][t][i]['sr'])
end = time.time()
print(">> %s task finished in %.0f secs " %
(green('Matching'), end - start))
return results
def eval_retrieval(descr, split): #WIP
print('>> Evaluating %s task' % green('retrieval'))
start = time.time()
q = pd.read_csv(
os.path.join(tskdir, 'retr_queries_split-' + split['name'] +
'.csv')).as_matrix()
d = pd.read_csv(
os.path.join(tskdir, 'retr_distractors_split-' + split['name'] +
'.csv')).as_matrix()
# q_std = np.std(q, axis=0)
# d_std = np.std(d, axis=0)
# print(q.shape)
# print(q_std.shape)
desc_q = descr_from_idx(descr, q).astype(np.float32)
desc_d = descr_from_idx(descr, d).astype(np.float32)
# distactor masking per sequence
m = dict((seq, d[:, 0] != seq) for seq in split['test'])
print(">> Please wait, computing distance matrix...")
D = dist_matrix(desc_q, desc_d, descr['distance'])
print(">> Distance matrix done.")
results = defaultdict(lambda: defaultdict(lambda: defaultdict(dict)))
N_distractors = desc_d.shape[0]
# at_ranks = [int(x*N_distractors) for x in [0.25,0.5,0.75,1]]
at_ranks = [100, 500, 1000, 5000, 10000, 15000, 20000]
pbar = tqdm(range(desc_q.shape[0]))
pbar.set_description("Processing retrieval task")
for i in pbar:
query_descr = desc_q[i]
for t in tp:
D_intra = get_query_intra_dists(descr, desc_q[i], q[i], t)
D_ = D[i, :]
D_ = D_[m[q[i][0]]]
gt = np.zeros_like(D_)
D_ = np.hstack((D_intra, D_))
# D_[0:5] = D_intra
gt = np.hstack((np.array([1, 1, 1, 1, 1]), gt))
# gt[0:5] = 1
for k in at_ranks:
pr, rc, ap = metrics.pr(-D_[0:k], gt[0:k])
# print (pr.shape,rc.shape)
# print ap
results[i][t][k]['ap'] = ap
# perm = np.argsort(D_[0:k], kind='mergesort',axis=0)
# gt_perm = gt[perm]
# mi_rank = np.mean(np.where(gt_perm))
# results[i][t][k]['mi_rank'] = mi_rank
end = time.time()
print(">> %s task finished in %.0f secs " %
(green('Retrieval'), end - start))
return results