def test(model, test_loader, losses, epoch, nr_query=Config.nr_query):
'''
return: cmc1, mAP
test model on testset and save result to log.
'''
val_start_time = time.time()
model.eval()
logger.info('testing', 'Start testing')
all_features, all_labels, all_cids = [], [], []
history = collections.defaultdict(list)
for i, (imgs, labels, cids, types,
colors) in tqdm(enumerate(test_loader),
desc='testing on epoch-{}'.format(epoch),
total=len(test_loader)):
imgs, labels, cids = imgs.cuda(), labels.cuda(), cids.cuda()
types, colors = types.cuda(), colors.cuda()
f_norm, p_type, p_color = model(imgs)
triplet_hard_loss = losses['triplet_hard_loss'](f_norm, labels)
acc_type = accuracy(p_type, types)[0]
acc_color = accuracy(p_color, colors)[0]
history['triplet_hard_loss'].append(float(triplet_hard_loss))
history['acc_type'].append(float(acc_type))
history['acc_color'].append(float(acc_color))
all_features.append(f_norm.cpu().detach().numpy())
all_labels.append(labels.cpu().detach().numpy())
all_cids.append(cids.cpu().detach().numpy())
all_features = np.concatenate(all_features, axis=0)
all_labels = np.concatenate(all_labels, axis=0)
all_cids = np.concatenate(all_cids, axis=0)
q_f, g_f = all_features[:nr_query], all_features[nr_query:]
q_ids, g_ids = all_labels[:nr_query], all_labels[nr_query:]
q_cids, g_cids = all_cids[:nr_query], all_cids[nr_query:]
print('Compute CMC and mAP')
distance_matrix = get_L2distance_matrix_numpy(q_f, g_f)
cmc, mAP = get_cmc_map(distance_matrix, q_ids, g_ids, q_cids, g_cids)
val_end_time = time.time()
time_spent = sec2min_sec(val_start_time, val_end_time)
text = 'Finish testing epoch {:>3}, time spent: [{:>3}mins{:>3}s], performance:\n##'.format(
epoch, time_spent[0], time_spent[1])
text += '|CMC1:{:>5.4f} |mAP:{:>5.4f}'.format(cmc[0], mAP)
for k, vlist in history.items():
v = float(mean(vlist))
text += '|{}:{:>5.4f} '.format(k, v)
logger.add_scalar('TEST/' + k, v, epoch)
logger.info('testing', text)
logger.add_scalar('TEST/cmc1', cmc[0], epoch)
logger.add_scalar('TEST/cmc5', cmc[4], epoch)
logger.add_scalar('TEST/cmc10', cmc[9], epoch)
logger.add_scalar('TEST/mAP', mAP, epoch)
return cmc, mAP
def test(model, test_loader, losses, epoch, nr_query=Config.nr_query):
'''
return: cmc1, mAP
test model on testset and save result to log.
'''
val_start_time = time.time()
model.eval()
logger.info('testing', 'Start testing')
all_features_gpu, all_labels, all_cids, all_mask_gpu = [], [], [], []
all_attn_dist = []
history = collections.defaultdict(list)
for i, (imgs, labels,
cids) in tqdm(enumerate(test_loader),
desc='testing on epoch-{}'.format(epoch),
total=len(test_loader)):
imgs, labels, cids = imgs.cuda(), labels.cuda(), cids.cuda()
f_norm, f_mask = model(imgs)
triplet_hard_loss, dist = losses['triplet_hard_loss'](f_norm, f_mask,
labels)
history['triplet_hard_loss'].append(float(triplet_hard_loss))
all_features_gpu.append(f_norm)
all_labels.append(labels.cpu().detach().numpy())
all_cids.append(cids.cpu().detach().numpy())
all_mask_gpu.append(f_mask)
all_attn_dist.append(dist.cpu().detach().numpy())
all_features_gpu = torch.cat(all_features_gpu, axis=0)
all_features = all_features_gpu.cpu().detach().numpy()
all_labels = np.concatenate(all_labels, axis=0)
all_cids = np.concatenate(all_cids, axis=0)
all_mask_gpu = torch.cat(all_mask_gpu, axis=0)
q_f_gpu, g_f_gpu = all_features_gpu[:nr_query], all_features_gpu[nr_query:]
q_f, g_f = all_features[:nr_query], all_features[nr_query:]
q_ids, g_ids = all_labels[:nr_query], all_labels[nr_query:]
q_cids, g_cids = all_cids[:nr_query], all_cids[nr_query:]
q_mask, g_mask = all_mask_gpu[:nr_query], all_mask_gpu[nr_query:]
print('Compute CMC and mAP')
attn_distance_matrix = get_L2distance_matrix_attn_batch(
q_f_gpu, g_f_gpu, q_mask, g_mask, temp=Config.temperature)
attn_distance_matrix = attn_distance_matrix.cpu().detach().numpy()
distance_matrix = get_L2distance_matrix_numpy(q_f, g_f)
cmc_a, mAP_a = get_cmc_map(attn_distance_matrix, q_ids, g_ids, q_cids,
g_cids)
cmc, mAP = get_cmc_map(distance_matrix, q_ids, g_ids, q_cids, g_cids)
val_end_time = time.time()
time_spent = sec2min_sec(val_start_time, val_end_time)
text = 'Finish testing epoch {:>3}, time spent: [{:>3}mins{:>3}s], performance:\n##'.format(
epoch, time_spent[0], time_spent[1])
text += 'With ATTENTION> |CMC1:{:>5.4f} |mAP:{:>5.4f} '.format(
cmc_a[0], mAP_a)
text += 'W/O attention> |CMC1:{:>5.4f} |mAP:{:>5.4f}'.format(cmc[0], mAP)
for k, vlist in history.items():
v = float(mean(vlist))
text += '|{}:{:>5.4f} '.format(k, v)
logger.add_scalar('TEST/' + k, v, epoch)
logger.info('testing', text)
logger.add_scalar('TEST/cmc1_a', cmc_a[0], epoch)
logger.add_scalar('TEST/cmc5_a', cmc_a[4], epoch)
logger.add_scalar('TEST/mAP_a', mAP_a, epoch)
logger.add_scalar('TEST/cmc1', cmc[0], epoch)
logger.add_scalar('TEST/cmc5', cmc[4], epoch)
logger.add_scalar('TEST/mAP', mAP, epoch)
return cmc_a, mAP_a, cmc, mAP
def test(model,
bclassifier,
test_loader,
losses,
epoch,
nr_query=Config.nr_query):
'''
return: cmc1, mAP
test model on testset and save result to log.
'''
val_start_time = time.time()
model.eval()
bclassifier.eval()
logger.info('testing', 'Start testing')
all_features_gpu, all_labels, all_cids, all_mask = [], [], [], []
for i, (imgs, labels, cids) in tqdm(enumerate(test_loader),
desc='extracting features',
total=len(test_loader)):
imgs, labels, cids = imgs.cuda(), labels.cuda(), cids.cuda()
f_norm, f_mask = model(imgs)
all_features_gpu.append(f_norm)
all_labels.append(labels)
all_mask.append(f_mask)
all_cids.append(cids)
features = torch.cat(all_features_gpu, axis=0)
alllabels = torch.cat(all_labels, axis=0)
allmask = torch.cat(all_mask, axis=0)
allcids = torch.cat(all_cids, axis=0)
q_f_gpu, g_f_gpu = features[:nr_query, ...], features[nr_query:, ...]
q_ids, g_ids = alllabels[:nr_query, ...], alllabels[nr_query:, ...]
q_cids, g_cids = allcids[:nr_query], allcids[nr_query:]
q_mask, g_mask = allmask[:nr_query], allmask[nr_query:]
pk_pros = []
for start in tqdm(range(0, q_f_gpu.shape[0], Config.eval_P),
desc='computing similarity'):
end = min(start + Config.eval_P, q_f_gpu.shape[0])
p_pros = []
for kstart in range(0, g_f_gpu.shape[0], Config.eval_K):
kend = min(kstart + Config.eval_K, g_f_gpu.shape[0])
pros = bclassifier(q_f_gpu[start:end, ...], g_f_gpu[kstart:kend,
...])
pros = pros.reshape(end - start, kend - kstart, 2)
p_pros.append(pros)
p_pros = torch.cat(p_pros, axis=1)
pk_pros.append(p_pros)
pros = torch.cat(pk_pros, axis=0)
pairlabels = idlabel2pairlabel(q_ids, g_ids)
acc = accuracy(pros.reshape(-1, 2), pairlabels)[0]
distance_matrix = pros[:, :, 0].cpu().detach().numpy()
q_ids = q_ids.cpu().detach().numpy()
g_ids = g_ids.cpu().detach().numpy()
q_cids = q_cids.cpu().detach().numpy()
g_cids = g_cids.cpu().detach().numpy()
print('Compute CMC and mAP')
cmc, mAP = get_cmc_map(distance_matrix, q_ids, g_ids, q_cids, g_cids)
val_end_time = time.time()
time_spent = sec2min_sec(val_start_time, val_end_time)
text = 'Finish testing epoch {:>3}, time spent: [{:>3}mins{:>3}s], performance:\n##'.format(
epoch, time_spent[0], time_spent[1])
text += 'W/O attention> |CMC1:{:>5.4f} |mAP:{:>5.4f} |ACC:{:>5.4f} '.format(
cmc[0], mAP, acc)
logger.info('testing', text)
logger.add_scalar('TEST/cmc1', cmc[0], epoch)
logger.add_scalar('TEST/cmc5', cmc[4], epoch)
logger.add_scalar('TEST/mAP', mAP, epoch)
logger.add_scalar('TEST/acc', acc, epoch)
return cmc, mAP, acc