def test(test_loader, model, epoch):
# switch to evaluate mode
model.eval()
labels, distances = [], []
pbar = tqdm(enumerate(test_loader))
for batch_idx, (data_a, data_p, label) in pbar:
if args.cuda:
data_a, data_p = data_a.cuda(), data_p.cuda()
data_a, data_p, label = Variable(data_a, volatile=True), \
Variable(data_p, volatile=True), Variable(label)
# compute output
out_a, out_p = model(data_a), model(data_p)
dists = l2_dist.forward(
out_a, out_p
) #torch.sqrt(torch.sum((out_a - out_p) ** 2, 1)) # euclidean distance
distances.append(dists.data.cpu().numpy())
labels.append(label.data.cpu().numpy())
if batch_idx % args.log_interval == 0:
pbar.set_description('Test Epoch: {} [{}/{} ({:.0f}%)]'.format(
epoch, batch_idx * len(data_a), len(test_loader.dataset),
100. * batch_idx / len(test_loader)))
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array(
[subdist[0] for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print('\33[91mTest set: Accuracy: {:.8f}\n\33[0m'.format(
np.mean(accuracy)))
logger.log_value('Test Accuracy', np.mean(accuracy))
plot_roc(fpr, tpr, figure_name="roc_test_epoch_{}.png".format(epoch))
def own_train(train_loader, model, triploss, optimizer, epoch, data_size):
model.train()
labels, distances = [], []
triplet_loss_sum = 0.0
for batch_idx, (data_a, data_p, data_n, label_p,
label_n) in enumerate(train_loader):
anc_img, pos_img, neg_img = data_a.to(device), data_p.to(
device), data_n.to(device)
with torch.set_grad_enabled(True):
anc_embed, pos_embed, neg_embed = model(anc_img), model(
pos_img), model(neg_img)
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist < args.margin).cpu().numpy().flatten()
hard_triplets = np.where(all == 1)
if len(hard_triplets) == 0:
continue
anc_hard_embed = anc_embed[hard_triplets]
pos_hard_embed = pos_embed[hard_triplets]
neg_hard_embed = neg_embed[hard_triplets]
anc_hard_img = anc_img[hard_triplets]
pos_hard_img = pos_img[hard_triplets]
neg_hard_img = neg_img[hard_triplets]
model.forward_classifier(anc_hard_img)
model.forward_classifier(pos_hard_img)
model.forward_classifier(neg_hard_img)
triplet_loss = triploss.forward(anc_hard_embed, pos_hard_embed,
neg_hard_embed)
logger.log_value('triplet_loss', triplet_loss)
optimizer.zero_grad()
triplet_loss.backward()
optimizer.step()
adjust_learning_rate(optimizer)
distances.append(pos_dist.data.cpu().numpy())
labels.append(np.ones(pos_dist.size(0)))
distances.append(neg_dist.data.cpu().numpy())
labels.append(np.zeros(neg_dist.size(0)))
triplet_loss_sum += triplet_loss.item()
avg_triplet_loss = triplet_loss_sum / data_size['train']
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array([subdist for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format('train',
avg_triplet_loss))
print(' {} set - Accuracy = {:.8f}'.format('train',
np.mean(accuracy)))
logger.log_value('Train Accuracy', np.mean(accuracy))
plot_roc(fpr, tpr, figure_name="roc_train_epoch_{}.png".format(epoch))
torch.save({
'epoch': epoch + 1,
'state_dict': model.state_dict()
}, '{}/checkpoint_{}.pth'.format(LOG_DIR, epoch))
def test(model, queryloader, galleryloader, use_gpu, ranks=[1, 5, 10, 20]):
batch_time = AverageMeter()
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
if use_gpu: imgs = imgs.cuda()
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".
format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
end = time.time()
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu: imgs = imgs.cuda()
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print("Extracted features for gallery set, obtained {}-by-{} matrix".
format(gf.size(0), gf.size(1)))
print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(
batch_time.avg, args.test_batch))
m, n = qf.size(0), gf.size(0)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat,
q_pids,
g_pids,
q_camids,
g_camids,
use_metric_cuhk03=args.use_metric_cuhk03)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def test(model,
queryloader,
galleryloader,
pool,
use_gpu,
ranks=[1, 5, 10, 20]):
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
if use_gpu: imgs = imgs.cuda()
b, s, c, h, w = imgs.size()
imgs = imgs.view(b * s, c, h, w)
features = model(imgs)
features = features.view(b, s, -1)
if pool == 'avg':
features = torch.mean(features, 1)
else:
features, _ = torch.max(features, 1)
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".
format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu: imgs = imgs.cuda()
b, s, c, h, w = imgs.size()
imgs = imgs.view(b * s, c, h, w)
features = model(imgs)
features = features.view(b, s, -1)
if pool == 'avg':
features = torch.mean(features, 1)
else:
features, _ = torch.max(features, 1)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print("Extracted features for gallery set, obtained {}-by-{} matrix".
format(gf.size(0), gf.size(1)))
print("Computing distance matrix")
m, n = qf.size(0), gf.size(0)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def test(model, queryloader, galleryloader, use_gpu, ranks=[1, 5, 10, 20]):
batch_time = AverageMeter()
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
if use_gpu: imgs = imgs.cuda()
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
end = time.time()
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu: imgs = imgs.cuda()
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print("Extracted features for gallery set, obtained {}-by-{} matrix".format(gf.size(0), gf.size(1)))
print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(batch_time.avg, args.test_batch))
m, n = qf.size(0), gf.size(0)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids, use_metric_cuhk03=args.use_metric_cuhk03)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r-1]))
print("------------------")
return cmc[0]
def test(model,
queryloader,
galleryloader,
pool,
use_gpu,
ranks=[1, 5, 10, 20]):
model.eval()
qf, q_pids, q_camids = [], [], []
features = []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
# print(batch_idx)
if use_gpu:
imgs = imgs.cuda()
with torch.no_grad():
imgs = Variable(imgs)
# b=1, n=number of clips, s=16
b, n, s, c, h, w = imgs.size()
assert (b == 1)
imgs = imgs.view(b * n, s, c, h, w)
# mid = int(b*n/2)
# l1 = imgs[0:mid]
# l2 = imgs[mid:]
# feature1 = model(l1)
# feature2 = model(l2)
if args.dataset == 'mars':
maxratio = 1 / 4
else:
maxratio = 3 / 4
for index, img in enumerate(imgs[0:int(b * n * maxratio)]):
img = img.unsqueeze(0)
feature = model(img)
# feature = feature.unsqeeze(0)
if index == 0:
features = feature
features = torch.cat((features, feature), 0)
# features = model(imgs)
# features = features.view(n, -1)
features = torch.mean(features, 0)
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.stack(qf)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".format(
qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
features = []
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu:
imgs = imgs.cuda()
with torch.no_grad():
imgs = Variable(imgs)
b, n, s, c, h, w = imgs.size()
imgs = imgs.view(b * n, s, c, h, w)
assert (b == 1)
features = model(imgs)
features = features.view(n, -1)
if pool == 'avg':
features = torch.mean(features, 0)
else:
features, _ = torch.max(features, 0)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.stack(gf)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print(
"Extracted features for gallery set, obtained {}-by-{} matrix".format(
gf.size(0), gf.size(1)))
print("Computing distance matrix")
m, n = qf.size(0), gf.size(0)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(
m, n) + torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
str(recall) + ' f1 ' + str(f1))
predictions = model.predict(x_features)
precision, recall, threshold, f1 = evaluation_on_balanced(
predictions, y_labels, is_speaking_labels, threshold)
print('Balanced test results: Prec ' + str(precision) + ' Recall ' +
str(recall) + ' f1 ' + str(f1))
print('Threshold ' + str(threshold))
else:
dev_preds = model.predict(x_dev)
precision, recall, threshold, f1 = eval_metrics.threshold_evaluate(
dev_preds, y_dev, False)
print('Results on Dev: Prec ' + str(precision) + ' Recall ' +
str(recall) + ' f1 ' + str(f1) + ' threshold ' + str(threshold))
predictions = model.predict(x_features)
precision, recall, f1 = eval_metrics.evaluate(predictions, y_labels,
threshold)
print('Results: Prec ' + str(precision) + ' Recall ' + str(recall) +
' f1 ' + str(f1))
precision, recall, f1 = eval_metrics.random_baseline(y_labels)
print('Random baseline: Prec ' + str(precision) + ' Recall ' +
str(recall) + ' f1 ' + str(f1))
precision, recall, f1 = eval_metrics.speech_only_baseline(
y_labels, is_speaking_labels)
print('Speech only baseline: Prec ' + str(precision) + ' Recall ' +
str(recall) + ' f1 ' + str(f1))
precision, recall, f1 = eval_metrics.content_words_baseline(
y_labels, is_speaking_labels, is_function_word_labels)
print('Content word baseline: Prec ' + str(precision) + ' Recall ' +
str(recall) + ' f1 ' + str(f1))
failure_analysis(y_labels, predictions, test_audio_files, test_times)
def test(model,
queryloader,
galleryloader,
use_gpu,
ranks=[1, 5, 10, 20],
use_salience=False,
use_parsing=False,
save_dir="",
epoch=-1,
save_rank=False,
use_re_ranking=False):
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
qimgs = []
for batch_idx, tuple_i in enumerate(queryloader):
if use_salience and not use_parsing:
imgs, pids, camids, salience_imgs, qimg = tuple_i
elif not use_salience and use_parsing:
imgs, pids, camids, parsing_imgs, qimg = tuple_i
elif use_salience and use_parsing:
imgs, pids, camids, salience_imgs, parsing_imgs, qimg = tuple_i
else:
imgs, pids, camids, qimg = tuple_i
if use_gpu:
imgs = imgs.cuda()
if use_salience and not use_parsing:
features = model(imgs, salience_masks=salience_imgs)
elif not use_salience and use_parsing:
features = model(imgs, parsing_masks=parsing_imgs)
elif use_salience and use_parsing:
features = model(imgs,
salience_masks=salience_imgs,
parsing_masks=parsing_imgs)
else:
features = model(imgs)
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qimgs.extend(qimg)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
qimgs = np.asarray(qimgs)
print("Extracted features for query set, obtained {}-by-{} matrix".
format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
gimgs = []
for batch_idx, tuple_i in enumerate(galleryloader):
if use_salience and not use_parsing:
imgs, pids, camids, salience_imgs, gimg = tuple_i
elif not use_salience and use_parsing:
imgs, pids, camids, parsing_imgs, gimg = tuple_i
elif use_salience and use_parsing:
imgs, pids, camids, salience_imgs, parsing_imgs, gimg = tuple_i
else:
imgs, pids, camids, gimg = tuple_i
if use_gpu:
imgs = imgs.cuda()
if use_salience and not use_parsing:
features = model(imgs, salience_masks=salience_imgs)
elif not use_salience and use_parsing:
features = model(imgs, parsing_masks=parsing_imgs)
elif use_salience and use_parsing:
features = model(imgs,
salience_masks=salience_imgs,
parsing_masks=parsing_imgs)
else:
features = model(imgs)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gimgs.extend(gimg)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
gimgs = np.asarray(gimgs)
print("Extracted features for gallery set, obtained {}-by-{} matrix".
format(gf.size(0), gf.size(1)))
print("Computing distance matrix")
if use_re_ranking:
qg_distmat = get_distance_matrix(qf, gf)
gg_distmat = get_distance_matrix(gf, gf)
qq_distmat = get_distance_matrix(qf, qf)
print("Re-ranking feature maps")
distmat = re_ranking(qg_distmat, qq_distmat, gg_distmat)
else:
distmat = get_distance_matrix(qf, gf)
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat,
q_pids,
g_pids,
q_camids,
g_camids,
use_metric_cuhk03=args.use_metric_cuhk03,
query_img_paths=qimgs,
gallery_img_paths=gimgs,
save_dir=save_dir,
epoch=epoch,
save_rank=save_rank)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def test(epoch, model, queryloader, galleryloader, use_gpu=True, ranks=[1, 5, 10, 20], summary=None):
batch_time = AverageMeter()
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
if use_gpu:
imgs = imgs.cuda()
end = time.time()
features = extract_feature(imgs,model)
# features = features/torch.norm(features,p=2,dim=1,keepdim=True)
batch_time.update(time.time() - end)
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu:
imgs = imgs.cuda()
end = time.time()
features = extract_feature(imgs,model)
# features = features/torch.norm(features,p=2,dim=1,keepdim=True)
batch_time.update(time.time() - end)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
if args.save_fea:
print("Saving result.mat to {}".format(args.save_dir))
result = {'gallery_f':gf.numpy(),'gallery_label':g_pids,'gallery_cam':g_camids,'query_f':qf.numpy(),'query_label':q_pids,'query_cam':q_camids}
sio.savemat(os.path.join(args.save_dir, 'dp_result.mat'),result)
return
print("Extracted features for gallery set, obtained {}-by-{} matrix".format(gf.size(0), gf.size(1)))
print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(batch_time.avg, args.test_batch))
m, n = qf.size(0), gf.size(0)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids, use_metric_cuhk03=args.use_metric_cuhk03)
print("----------Results-----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r-1]))
#print("---- Start Reranking... ----")
#rerank_cmc, rerank_mAP = rerank_main(qf,q_camids,q_pids,gf,g_camids,g_pids)
#print("Rerank_mAP: {:.1%}".format(rerank_mAP))
#print("Rerank_CMC curve")
#for r in ranks:
# print("Rank-{:<3}: {:.1%}".format(r, rerank_cmc[r-1]))
print("----------------------------")
if summary is not None:
summary.add_scalars('rank result', {'rank1': round(cmc[0],3) * 100, 'rank5': round(cmc[1],3) * 100, 'mAP': round(mAP,3) * 100}, epoch)
return cmc[0] #rerank_cmc[0]
# compute output
out_a, out_p = model(data_a), model(data_p)
dists = l2_dist.forward(out_a,out_p)#torch.sqrt(torch.sum((out_a - out_p) ** 2, 1)) # euclidean distance
distances.append(dists.data.cpu().numpy())
labels.append(label.data.cpu().numpy())
if batch_idx % args.log_interval == 0:
pbar.set_description('Test Epoch: {} [{}/{} ({:.0f}%)]'.format(
epoch, batch_idx * len(data_a), len(test_loader.dataset),
100. * batch_idx / len(test_loader)))
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array([subdist[0] for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances,labels)
print('\33[91mTest set: Accuracy: {:.8f}\n\33[0m'.format(np.mean(accuracy)))
logger.log_value('Test Accuracy', np.mean(accuracy))
plot_roc(fpr,tpr,figure_name="roc_test_epoch_{}.png".format(epoch))
def testaccuracy(test_loader,model,epoch):
# switch to evaluate mode
model.eval()
pbar = tqdm(enumerate(test_loader))
prec = []
correct = 0
for batch_idx, (data, label) in pbar:
data_v = Variable(data.cuda())
target_value = Variable(label)
def test_gcn_person_batch(model,
queryloader,
querygalleryloader,
galleryloader,
pool,
use_gpu,
ranks=[1, 5, 10, 20]):
model.eval()
g_bs = 16
q_pids, q_pids_p, q_camids, q_camids_p = [], [], [], []
g_pids, g_pids_p, g_camids, g_camids_p = [], [], [], []
for batch_idx, (_, gids, pimgs, pids, camids, _) in enumerate(queryloader):
q_pids.extend(gids)
q_pids_p.extend(pids)
q_camids.extend(camids)
q_camids_p.extend([camids] * len(pids))
#print(camids)
q_pids = np.asarray(q_pids)
q_pids_p = np.asarray(q_pids_p)
q_pids_p = np.squeeze(q_pids_p)
q_camids = np.asarray(q_camids)
q_camids_p = np.asarray(q_camids_p)
max_qcam = camids + 1
print(q_pids.shape, q_pids_p.shape, q_camids.shape, q_camids_p.shape)
for batch_idx, (_, gids, pimgs, pids, camids,
_) in enumerate(querygalleryloader):
g_pids.extend(gids)
#print(gids, pids, camids)
tmp_pids = []
for j in range(g_bs):
tmp_pids.append([])
for i in range(len(pids)):
tmp_pids[j].append(pids[i][j])
# tmp_pids -> list g_bs * 5
for i in range(g_bs):
g_pids_p.extend(tmp_pids[i])
#print(camids)
#print(camids[i].item())
g_camids.extend([camids[i]])
g_camids_p.extend([camids[i]] * len(tmp_pids[i]))
#g_camids_p.extend([camids]* len(pids))
#g_pids = np.asarray(g_pids)
#g_pids_p = np.asarray(g_pids_p)
#g_camids = np.asarray(g_camids)
#g_camids_p = np.asarray(g_camids_p)
#print(g_pids.shape, g_pids_p.shape, g_camids.shape, g_camids_p.shape)
for batch_idx, (_, gids, pimgs, pids, camids,
_) in enumerate(galleryloader):
g_pids.extend(gids)
#print(gids, pids, camids)
tmp_pids = []
for j in range(g_bs):
tmp_pids.append([])
for i in range(len(pids)):
tmp_pids[j].append(pids[i][j])
# tmp_pids -> list g_bs * 5
for i in range(g_bs):
g_pids_p.extend(tmp_pids[i])
#print(camids)
#print(camids[i].item())
g_camids.extend([camids[i]])
g_camids_p.extend([camids[i] + max_qcam] * len(tmp_pids[i]))
#g_camids_p.extend([camids]* len(pids))
g_pids = np.asarray(g_pids)
g_pids_p = np.asarray(g_pids_p)
g_camids = np.asarray(g_camids)
g_camids_p = np.asarray(g_camids_p)
print(g_pids.shape, g_pids_p.shape, g_camids.shape, g_camids_p.shape)
m, n = q_pids.shape[0], g_pids.shape[0]
distmat = torch.zeros((m, n))
m, n = q_pids_p.shape[0], g_pids_p.shape[0]
distmat_p = torch.zeros((m, n))
p_start = 0
p_end = 0
with torch.no_grad():
for batch_idx, (_, gids, pimgs, pids, camids,
lenp) in enumerate(queryloader):
#if batch_idx < 1720:
# continue
start_time = time.time()
if use_gpu:
pimgs = pimgs.cuda()
pimgs = Variable(pimgs)
# b=1, n=number of clips, s=16
b, s, c, h, w = pimgs.size()
#pimgs = pimgs.permute(1, 0, 2, 3, 4)
assert (b == 1)
pimgs = pimgs.repeat(g_bs, 1, 1, 1, 1)
pimgs = pimgs.view(g_bs * s, c, h, w)
#pimgs = pimgs.view(s, c, h, w)
num_nodes = s
adj = []
adj0 = torch.ones((lenp, lenp))
if use_gpu:
adj0 = adj0.cuda()
adj0 = Variable(adj0)
adj0.requires_gradient = False
for aa in range(g_bs):
adj.append(adj0)
p_start = batch_idx * s
p_end = (batch_idx + 1) * s
#print(p_start, p_end)
#print(batch_idx, g_bs, s)
g_start = 0
g_end = 0
for batch_idx_g, (_, gids_g, pimgs_g, pids_g, camids_g,
lenp_g) in enumerate(querygalleryloader):
if use_gpu:
pimgs_g = pimgs_g.cuda()
pimgs_g = Variable(pimgs_g)
# pimgs = pimgs.permute(1, 0, 2, 3, 4)
b, s, c, h, w = pimgs_g.size()
pimgs_g = pimgs_g.view(b * s, c, h, w)
#pimgs_g = pimgs_g.view(s, c, h, w)
assert (b == g_bs)
num_nodes = s
adj_g = []
for aa in range(g_bs):
adj1 = torch.ones((lenp_g[aa], lenp_g[aa]))
if use_gpu:
adj1 = adj1.cuda()
adj1 = Variable(adj1)
adj1.requires_gradient = False
adj_g.append(adj1)
features1, features2, features_p1, features_p2 = model(
pimgs, pimgs_g, adj, adj_g)
#print(features_p1[0].shape, features_p2[0].shape)
features_p1 = torch.cat(features_p1, dim=1)
features_p2 = torch.cat(features_p2, dim=1)
#print(features_p1.shape)
dist_p = torch.pow(features_p1, 2).sum(dim=1, keepdim=True) + \
torch.pow(features_p2, 2).sum(dim=1, keepdim=True).t()
dist_p.addmm_(1, -2, features_p1, features_p2.t())
#p_end = p_start + dist_p.shape[0]
#assert (p_end - p_start) == dist_p.shape[0]
#print(p_end-p_start, dist_p.shape[0])
g_end = g_start + dist_p.shape[1]
#print(dist_p.shape)
#print(features_p1.shape, features_p2.shape)
#print(distmat_p[p_start:p_end, g_start:g_end].shape)
#distmat_p[p_start:p_end, g_start:g_end] = dist_p
for i in range(g_bs):
distmat_p[p_start:p_end, g_start + i * s:g_start +
(i + 1) * s] = dist_p[i * s:(i + 1) * s,
i * s:(i + 1) * s]
#distmat_p[p_start:p_end, g_start:g_end] = dist_p
assert (g_end == g_start + (i + 1) * s)
g_start = g_end
#print(dist)
dist = F.pairwise_distance(features1, features2)
#print(dist.shape)
distmat[batch_idx,
batch_idx_g * g_bs:(batch_idx_g + 1) * g_bs] = dist
#distmat[batch_idx, batch_idx_g] = dist
#print(g_end)
max_batch_idx_g = batch_idx_g + 1
for batch_idx_g, (_, gids_g, pimgs_g, pids_g, camids_g,
lenp_g) in enumerate(galleryloader):
if use_gpu:
pimgs_g = pimgs_g.cuda()
pimgs_g = Variable(pimgs_g)
# pimgs = pimgs.permute(1, 0, 2, 3, 4)
b, s, c, h, w = pimgs_g.size()
pimgs_g = pimgs_g.view(b * s, c, h, w)
#pimgs_g = pimgs_g.view(s, c, h, w)
assert (b == g_bs)
num_nodes = s
adj_g = []
for aa in range(g_bs):
adj1 = torch.ones((lenp_g[aa], lenp_g[aa]))
if use_gpu:
adj1 = adj1.cuda()
adj1 = Variable(adj1)
adj1.requires_gradient = False
adj_g.append(adj1)
features1, features2, features_p1, features_p2 = model(
pimgs, pimgs_g, adj, adj_g)
#print(features_p1[0].shape, features_p2[0].shape)
features_p1 = torch.cat(features_p1, dim=1)
features_p2 = torch.cat(features_p2, dim=1)
#print(features_p1.shape)
dist_p = torch.pow(features_p1, 2).sum(dim=1, keepdim=True) + \
torch.pow(features_p2, 2).sum(dim=1, keepdim=True).t()
dist_p.addmm_(1, -2, features_p1, features_p2.t())
#p_end = p_start + dist_p.shape[0]
#assert (p_end - p_start) == dist_p.shape[0]
#print(p_end-p_start, dist_p.shape[0])
g_end = g_start + dist_p.shape[1]
#print(dist_p.shape)
#print(features_p1.shape, features_p2.shape)
#print(distmat_p[p_start:p_end, g_start:g_end].shape)
#distmat_p[p_start:p_end, g_start:g_end] = dist_p
for i in range(g_bs):
distmat_p[p_start:p_end, g_start + i * s:g_start +
(i + 1) * s] = dist_p[i * s:(i + 1) * s,
i * s:(i + 1) * s]
#distmat_p[p_start:p_end, g_start:g_end] = dist_p
assert (g_end == g_start + (i + 1) * s)
g_start = g_end
#print(dist)
dist = F.pairwise_distance(features1, features2)
#print(dist.shape)
distmat[batch_idx, (max_batch_idx_g + batch_idx_g) *
g_bs:(max_batch_idx_g + batch_idx_g + 1) * g_bs] = dist
#print(g_end)
#p_start = p_end
#print(batch_idx)
end_time = time.time()
print("image {:04d}, time : {:f}".format(batch_idx,
end_time - start_time))
distmat = distmat.numpy()
distmat_p = distmat_p.numpy()
#print(distmat)
print("Computing CMC and mAP")
print(distmat.shape, q_pids.shape, g_pids.shape, q_camids.shape,
g_camids.shape)
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
#cmc_p, mAP_p = evaluate_person(distmat_p, q_pids_p, g_pids_p, q_camids_p, g_camids_p)
print("Group Reid Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
print(distmat_p.shape, q_pids_p.shape, g_pids_p.shape, q_camids_p.shape,
g_camids_p.shape)
cmc_p, mAP_p = evaluate_person(distmat_p, q_pids_p, g_pids_p, q_camids_p,
g_camids_p)
print("Person Reid Results ----------")
print("mAP: {:.1%}".format(mAP_p))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc_p[r - 1]))
print("------------------")
return cmc[0]
def test_gcn(model,
queryloader,
galleryloader,
pool,
use_gpu,
ranks=[1, 5, 10, 20]):
model.eval()
q_pids, q_camids = [], []
g_pids, g_camids = [], []
for batch_idx, (_, gids, pimgs, pids, camids) in enumerate(queryloader):
q_pids.extend(gids)
q_camids.extend(camids)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
max_qcam = camids + 1
for batch_idx, (_, gids, pimgs, pids, camids) in enumerate(galleryloader):
g_pids.extend(gids)
g_camids.extend(camids + max_qcam)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
m, n = q_pids.shape[0], g_pids.shape[0]
distmat = torch.zeros((m, m + n))
g_camids = np.concatenate((q_camids, g_camids), axis=0)
g_pids = np.concatenate((q_pids, g_pids), axis=0)
with torch.no_grad():
for batch_idx, (_, gids, pimgs, pids,
camids) in enumerate(queryloader):
if use_gpu:
pimgs = pimgs.cuda()
pimgs = Variable(pimgs)
# b=1, n=number of clips, s=16
b, s, c, h, w = pimgs.size()
#pimgs = pimgs.permute(1, 0, 2, 3, 4)
assert (b == 1)
pimgs = pimgs.view(s, c, h, w)
num_nodes = s
adj = torch.ones((num_nodes, num_nodes))
if use_gpu:
adj = adj.cuda()
adj = Variable(adj)
for batch_idx_q, (_, gids_q, pimgs_q, pids_q,
camids_q) in enumerate(queryloader):
if use_gpu:
pimgs_q = pimgs_q.cuda()
pimgs_q = Variable(pimgs_q)
# pimgs = pimgs.permute(1, 0, 2, 3, 4)
b, s, c, h, w = pimgs_q.size()
pimgs_q = pimgs_q.view(s, c, h, w)
assert (b == 1)
num_nodes = s
adj_q = torch.ones((num_nodes, num_nodes))
if use_gpu:
adj_q = adj_q.cuda()
adj_q = Variable(adj_q)
features1, features2 = model(pimgs, pimgs_q, [adj], [adj_q])
#dist = torch.pow(features1, 2).sum(dim=1, keepdim=True) + \
# torch.pow(features2, 2).sum(dim=1, keepdim=True).t()
#dist.addmm_(1, -2, features1, features2.t())
#print(dist)
dist = F.pairwise_distance(features1, features2)
#print(dist)
distmat[batch_idx, batch_idx_q] = dist
for batch_idx_g, (_, gids_g, pimgs_g, pids_g,
camids_g) in enumerate(galleryloader):
if use_gpu:
pimgs_g = pimgs_g.cuda()
pimgs_g = Variable(pimgs_g)
# pimgs = pimgs.permute(1, 0, 2, 3, 4)
b, s, c, h, w = pimgs_g.size()
pimgs_g = pimgs_g.view(s, c, h, w)
assert (b == 1)
num_nodes = s
adj_g = torch.ones((num_nodes, num_nodes))
if use_gpu:
adj_g = adj_g.cuda()
adj_g = Variable(adj_g)
features1, features2 = model(pimgs, pimgs_g, [adj], [adj_g])
#dist = torch.pow(features1, 2).sum(dim=1, keepdim=True) + \
# torch.pow(features2, 2).sum(dim=1, keepdim=True).t()
#dist.addmm_(1, -2, features1, features2.t())
#print(dist)
dist = F.pairwise_distance(features1, features2)
#print(dist)
distmat[batch_idx, batch_idx_g + m] = dist
distmat = distmat.numpy()
#print(distmat)
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
'''
dist_qq = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, m) + \
torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, m).t()
dist_qq.addmm_(1, -2, qf, qf.t())
dist_qq = dist_qq.numpy()
dist_gg = torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, n).t()
dist_gg.addmm_(1, -2, gf, gf.t())
dist_gg = dist_gg.numpy()
dist_re_rank = re_ranking(distmat, dist_qq, dist_gg)
print("Computing CMC and mAP")
cmc, mAP = evaluate(dist_re_rank, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r-1]))
print("------------------")
'''
return cmc[0]
def main(args=None):
# parse arguments
if args is None:
args = sys.argv[1:]
args = parse_args(args)
# data generator
dataset = data_manager.init_imgreid_dataset(
dataset_path=args.dataset_path, name=args.dataset
)
query_generator = DataGenerator(dataset.query[0],
dataset.query[1],
camids=dataset.query[2],
batch_size=args.batch_size,
num_classes=dataset.num_query_pids,
target_size=(args.height, args.width),
learn_region=args.learn_region,
shuffle=True,
mode='inference')
gallery_generator = DataGenerator( dataset.gallery[0],
dataset.gallery[1],
camids=dataset.gallery[2],
batch_size=args.batch_size,
num_classes=dataset.num_gallery_pids,
target_size=(args.height, args.width),
learn_region=args.learn_region,
shuffle=True,
mode='inference')
model = modellib.HACNN(mode='inference',
num_classes=dataset.num_query_pids,
batch_size=args.batch_size,
learn_region=args.learn_region).model
load_weights(model, filepath=args.snapshot, by_name=True)
'''
img_path = '/Users/luke/Documents/ml_datasets/person_re_id/videotag_scene/dataset_7_lite/bounding_box_train/3434_0096.jpg'
img = image.load_img(img_path, target_size=(args.height, args.width))
img = image.img_to_array(img)
ouput = model.predict(np.array([img]) ,verbose=1)
print('ouput', np.array(ouput).shape)
print('ouput', np.argmax(ouput[0]), np.argmax(ouput[1]))
'''
# evaluate
qf, q_pids, q_camids = [], [], []
for index in range(len(query_generator)):
imgs, pids, camids = query_generator[index]
# print('query', index)
features = model.predict(imgs ,verbose=0)
# print('features', features)
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = np.concatenate(qf, axis=0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
# print(qf.shape)
gf, g_pids, g_camids = [], [], []
for index in range(len(gallery_generator)):
# print('gallery', index)
imgs, pids, camids = gallery_generator[index]
features = model.predict(imgs ,verbose=0)
# print('features', features)
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = np.concatenate(gf, axis=0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
m, n = qf.shape[0], gf.shape[0]
'''
# qf = qf*0.001
qf_pow = np.square(qf)
qf_sum = np.sum(qf_pow, axis=1, keepdims=True)
qf_ext = np.repeat(qf_sum, n, axis=1)
# gf = gf*0.01
gf_pow = np.square(gf)
gf_sum = np.sum(gf_pow, axis=1, keepdims=True)
gf_ext = np.repeat(gf_sum, m, axis=1)
gf_ext_t = gf_ext.T
distmat = qf_ext + gf_ext_t
distmat = distmat + np.dot(qf, gf.T)*(-2.0)
'''
print("Compute pairwise euclidean distances")
qf = np.expand_dims(qf, axis=1)
qf = np.repeat(qf, n, axis=1)
gf = np.expand_dims(gf, axis=0)
gf = np.repeat(gf, m, axis=0)
distmat = np.linalg.norm(qf-gf, axis=2, keepdims=True)
distmat = np.squeeze(distmat, axis=2)
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids, dataset_type=args.dataset)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
ranks=[1, 5, 10, 20]
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r-1]))
correct_matched = (g_pids == q_pids[:, np.newaxis]).astype(np.int32)
correct_dist_aver = np.multiply(distmat, correct_matched)
wrong_matched = (g_pids != q_pids[:, np.newaxis]).astype(np.int32)
wrong_dist_aver = np.multiply(distmat, wrong_matched)
print('정답의 distance 평균 : ', np.average(correct_dist_aver))
print('오답의 distance 평균 : ', np.average(wrong_dist_aver))
print("------------------")
import numpy as np from IPython import embed from eval_metrics import evaluate from eval_metrics import evaluate, plot_roc, plot_acc targets = np.random.randint(1, 10, (100)) distances = np.random.randn(100) tpr, fpr, acc, val, val_std, far = evaluate(distances, targets) embed() plot_roc(fpr, tpr, figure_name='./test.png')
def train_valid(model, optimizer, scheduler, epoch, dataloaders, data_size):
for phase in ['train', 'valid']:
labels, distances = [], []
triplet_loss_sum = 0.0
if phase == 'train':
scheduler.step()
model.train()
else:
model.eval()
for batch_idx, batch_sample in enumerate(dataloaders[phase]):
#break
anc_img = batch_sample['anc_img'].to(device)
pos_img = batch_sample['pos_img'].to(device)
neg_img = batch_sample['neg_img'].to(device)
pos_cls = batch_sample['pos_class'].to(device)
neg_cls = batch_sample['neg_class'].to(device)
with torch.set_grad_enabled(phase == 'train'):
# anc_embed, pos_embed and neg_embed are encoding(embedding) of image
anc_embed, pos_embed, neg_embed = model(anc_img), model(
pos_img), model(neg_img)
# choose the hard negatives only for "training"
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist <
args.margin).cpu().numpy().flatten()
if phase == 'train':
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
else:
hard_triplets = np.where(all >= 0)
anc_hard_embed = anc_embed[hard_triplets].to(device)
pos_hard_embed = pos_embed[hard_triplets].to(device)
neg_hard_embed = neg_embed[hard_triplets].to(device)
anc_hard_img = anc_img[hard_triplets].to(device)
pos_hard_img = pos_img[hard_triplets].to(device)
neg_hard_img = neg_img[hard_triplets].to(device)
pos_hard_cls = pos_cls[hard_triplets].to(device)
neg_hard_cls = neg_cls[hard_triplets].to(device)
anc_img_pred = model.forward_classifier(anc_hard_img).to(
device)
pos_img_pred = model.forward_classifier(pos_hard_img).to(
device)
neg_img_pred = model.forward_classifier(neg_hard_img).to(
device)
triplet_loss = TripletLoss(args.margin).forward(
anc_hard_embed, pos_hard_embed, neg_hard_embed).to(device)
if phase == 'train':
optimizer.zero_grad()
triplet_loss.backward()
optimizer.step()
dists = l2_dist.forward(anc_embed, pos_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
dists = l2_dist.forward(anc_embed, neg_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
triplet_loss_sum += triplet_loss.item()
avg_triplet_loss = triplet_loss_sum / data_size[phase]
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array(
[subdist for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format(
phase, avg_triplet_loss))
print(' {} set - Accuracy = {:.8f}'.format(
phase, np.mean(accuracy)))
with open('./log/{}_log_epoch{}.txt'.format(phase, epoch), 'w') as f:
f.write(
str(epoch) + '\t' + str(np.mean(accuracy)) + '\t' +
str(avg_triplet_loss))
if phase == 'train':
torch.save({
'epoch': epoch,
'state_dict': model.state_dict()
}, './log/checkpoint_epoch{}.pth'.format(epoch))
else:
plot_roc(fpr,
tpr,
figure_name='./log/roc_valid_epoch_{}.png'.format(epoch))
def evaluation(model,
args,
queryloader,
galleryloader,
use_gpu,
ranks=[1, 5, 10, 20]):
since = time.time()
model.eval()
qf, q_pids, q_camids = [], [], []
for batch_idx, (vids, pids, camids) in enumerate(queryloader):
if (batch_idx + 1) % 1000 == 0:
print("{}/{}".format(batch_idx + 1, len(queryloader)))
qf.append(extract(model, args, vids, use_gpu).squeeze())
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.stack(qf)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {} matrix".format(
qf.shape))
gf, g_pids, g_camids = [], [], []
for batch_idx, (vids, pids, camids) in enumerate(galleryloader):
if (batch_idx + 1) % 1000 == 0:
print("{}/{}".format(batch_idx + 1, len(galleryloader)))
gf.append(extract(model, args, vids, use_gpu).squeeze())
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.stack(gf)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
if 'mars' in args.dataset:
print('process the dataset mars!')
# gallery set must contain query set, otherwise 140 query imgs will not have ground truth.
gf = torch.cat((qf, gf), 0)
g_pids = np.append(q_pids, g_pids)
g_camids = np.append(q_camids, g_camids)
print("Extracted features for gallery set, obtained {} matrix".format(
gf.shape))
time_elapsed = time.time() - since
print('Extracting features complete in {:.0f}m {:.0f}s'.format(
time_elapsed // 60, time_elapsed % 60))
print("Computing distance matrix")
m, n = qf.size(0), gf.size(0)
distmat = torch.zeros((m, n))
distmat = -torch.mm(qf, gf.t())
distmat = distmat.data.cpu()
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.2%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.2%}".format(r, cmc[r - 1]))
print("------------------")
elapsed = round(time.time() - since)
elapsed = str(datetime.timedelta(seconds=elapsed))
print("Finished. Total elapsed time (h:m:s): {}.".format(elapsed))
return cmc[0]
def test(model, queryloader, galleryloader, use_gpu, ranks=[1, 5, 10, 20]):
since = time.time()
batch_time = AverageMeter()
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
end = time.time()
if use_gpu:
imgs = imgs.cuda()
n, c, h, w = imgs.size()
features = torch.FloatTensor(n, model.module.feat_dim).zero_()
for i in range(2):
if (i == 1):
imgs = fliplr(imgs, use_gpu)
f = model(imgs)
f = f.data.cpu()
features = features + f
batch_time.update(time.time() - end)
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".
format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
end = time.time()
if use_gpu:
imgs = imgs.cuda()
n, c, h, w = imgs.size()
features = torch.FloatTensor(n, model.module.feat_dim).zero_()
for i in range(2):
if (i == 1):
imgs = fliplr(imgs, use_gpu)
f = model(imgs)
f = f.data.cpu()
features = features + f
batch_time.update(time.time() - end)
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print("Extracted features for gallery set, obtained {}-by-{} matrix".
format(gf.size(0), gf.size(1)))
print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(
batch_time.avg, args.test_batch))
m, n = qf.size(0), gf.size(0)
distmat = torch.zeros((m, n))
if args.distance == 'euclidean':
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
else:
q_norm = torch.norm(qf, p=2, dim=1, keepdim=True)
g_norm = torch.norm(gf, p=2, dim=1, keepdim=True)
qf = qf.div(q_norm.expand_as(qf))
gf = gf.div(g_norm.expand_as(gf))
distmat = -torch.mm(qf, gf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def main():
parser = argparse.ArgumentParser(description="Tensorflow Contrastive Convolution")
parser.add_argument('--num_classes', default=10575, type=int,
metavar='N', help='number of classes (default: 5749)')
parser.add_argument('--iters', type=int, default = 200000, metavar='N',
help='number of iterations to train (default: 10)')
args = parser.parse_args()
dataset = CasiaFaceDataset()
testset = LFWDataset()
base_model = ConstractiveFourLayers()
gen_model = GenModel(512)
reg_model = Regressor(350*32)
idreg_model = IdentityRegressor(14*512*3*3, args.num_classes)
input1 = tf.placeholder(tf.float32, [None, 128, 128, 1])
input2 = tf.placeholder(tf.float32, [None, 128, 128, 1])
target = tf.placeholder(tf.float32, [None, 1])
c1 = tf.placeholder(tf.float32, [None, args.num_classes])
c2 = tf.placeholder(tf.float32, [None, args.num_classes])
A_list, B_list, org_kernel_1, org_kernel_2 = compute_contrastive_features(input1, input2, base_model, gen_model)
reg_1 = reg_model.forward(A_list)
reg_2 = reg_model.forward(B_list)
SAB = tf.add(reg_1, reg_2) / 2.0
hk1 = idreg_model.forward(org_kernel_1)
hk2 = idreg_model.forward(org_kernel_2)
# print("target", target)
# print("SAB", SAB)
loss1 = tf.losses.sigmoid_cross_entropy(multi_class_labels=target, logits=SAB)
cross_entropy_1 = tf.reduce_mean(-tf.reduce_sum(c1 * tf.log(hk1), reduction_indices=[1]))
cross_entropy_2 = tf.reduce_mean(-tf.reduce_sum(c2 * tf.log(hk2), reduction_indices=[1]))
loss2 = tf.add(cross_entropy_1, cross_entropy_2) * 0.5
loss = tf.add(loss1, loss2)
optimizer = tf.train.GradientDescentOptimizer(0.001).minimize(loss)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
f = open("result.txt", "w")
for iteration in range(args.iters):
data_1_batch, data_2_batch, c1_batch, c2_batch, target_batch = dataset.get_batch(batch_size=GLOBAL_BATCH_SIZE)
# print(target_batch.shape)
# data_1_cur, data_2_cur, c1_cur, c2_cur, target_cur = sess.run([data_1_batch, data_2_batch, c1_batch, c2_batch, target_batch])
_, loss_val, loss1_val, loss2_val, reg1_val, reg2_val = sess.run([optimizer, loss, loss1, loss2, reg_1, reg_2],
feed_dict={input1: data_1_batch, input2: data_2_batch, c1: c1_batch, c2: c2_batch, target: target_batch})
print("Itera {0} : loss = {1}, loss1 = {2}, loss2 = {3}".format(iteration, loss_val, loss1_val, loss2_val))
f.write("Itera {0} : loss = {1}, loss1 = {2}, loss2 = {3}\r\n".format(iteration, loss_val, loss1_val, loss2_val))
f.flush()
if(iteration != 0 and iteration % 100 == 0):
acc_pool, start_time = [], time.time()
for i in range(50):
test_1_batch, test_2_batch, label_batch = testset.get_batch(batch_size=GLOBAL_BATCH_SIZE)
# test_1_cur, test_2_cur, label_cur = sess.run([data_1_batch, data_2_batch, label_batch])
# out1_a, out1_b, k1, k2 = sess.run(compute_contrastive_features(test_1_batch, test_2_batch, base_model, gen_model))
SAB_val, reg1_val, reg2_val = sess.run([SAB, reg_1, reg_2], feed_dict={input1: test_1_batch, input2: test_2_batch})
# print("SAB", SAB_val)
# print("1v", reg1_val)
# print("2v", reg2_val)
dists = np.array(SAB_val).reshape((-1, 1))
# print(dists)
labels = np.array(label_batch).reshape((-1, 1))
# print(labels)
accuracy = evaluate(1.0 - dists, labels)
acc_pool.append(np.mean(accuracy))
print("Acc(%.2f)"%(time.time()-start_time), np.mean(acc_pool), acc_pool)
f.write("Acc" + str(np.mean(acc_pool)) + str(acc_pool) + str("\r\n"))
f.flush()
f.close()
def test(self, queryloader, galleryloader, args, ranks=[1, 5, 10, 20]):
self.brain.eval()
self.feature_extractor.eval()
buffer_file = '../scratch/' + args.dataset + '_test_features.pth'
if not os.path.exists(buffer_file):
# if the buffer files saved with features already existing, load buffer file
# if not, extract the features from feature extractor
qindividualf, qmeanf, q_pids, q_camids = [], [], [], []
print('extracting query feats')
for batch_idx, (imgs, pids,
camids) in enumerate(tqdm(queryloader)):
if not args.use_cpu:
imgs = imgs.cuda()
with torch.no_grad():
# b=1, n=number of clips, s=16
b, n, s, c, h, w = imgs.size()
assert (b == 1)
imgs = imgs.view(b * n, s, c, h, w)
individual_features = get_features(self.feature_extractor,
imgs,
args.test_num_tracks)
mean_features = torch.mean(individual_features, 0)
individual_features = individual_features.data.cpu()
mean_features = mean_features.data.cpu()
qindividualf.append(individual_features)
qmeanf.append(mean_features)
q_pids.extend(pids)
q_camids.extend(camids)
torch.cuda.empty_cache()
qmeanf = torch.stack(qmeanf)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".
format(qmeanf.size(0), qmeanf.size(1)))
gindividualf, gmeanf, g_pids, g_camids = [], [], [], []
print('extracting gallery feats')
for batch_idx, (imgs, pids,
camids) in enumerate(tqdm(galleryloader)):
if not args.use_cpu:
imgs = imgs.cuda()
with torch.no_grad():
b, n, s, c, h, w = imgs.size()
imgs = imgs.view(b * n, s, c, h, w)
assert (b == 1)
# handle chunked data
individual_features = get_features(self.feature_extractor,
imgs,
args.test_num_tracks)
mean_features = torch.mean(individual_features, 0)
torch.cuda.empty_cache()
individual_features = individual_features.data.cpu()
mean_features = mean_features.data.cpu()
gindividualf.append(individual_features)
gmeanf.append(mean_features)
g_pids.extend(pids)
g_camids.extend(camids)
gmeanf = torch.stack(gmeanf)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print(
"Extracted features for gallery set, obtained {}-by-{} matrix".
format(gmeanf.size(0), gmeanf.size(1)))
torch.save(
{
'query': {
'meanft': qmeanf,
'individualft': qindividualf,
'pids': q_pids,
'camids': q_camids
},
'gallery': {
'meanft': gmeanf,
'individualft': gindividualf,
'pids': g_pids,
'camids': g_camids
}
}, buffer_file)
else:
# load the buffer file
print('loading and extraction information/features from file',
buffer_file)
buffer = torch.load(buffer_file)
qmeanf = buffer['query']['meanft']
qindividualf = buffer['query']['individualft']
q_camids = buffer['query']['camids']
q_pids = buffer['query']['pids']
gmeanf = buffer['gallery']['meanft']
gindividualf = buffer['gallery']['individualft']
g_camids = buffer['gallery']['camids']
g_pids = buffer['gallery']['pids']
print("Computing distance matrix for allframes evaluation (baseline)")
m, n = qmeanf.size(0), gmeanf.size(0)
distmat = torch.pow(qmeanf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gmeanf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qmeanf, gmeanf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
print('Computing distance matrix from DQN network')
distmat = torch.zeros(m, n)
instance_rewards = torch.zeros(m, n)
comparisons = torch.zeros(m, n)
print(m, n)
for i, qid in tqdm(enumerate(q_pids)):
q_features = self.select_random_features(qindividualf[i],
args.rl_seq_len)
for j, gid in enumerate(g_pids):
# print(qindividualf[i].shape, gindividualf[j].shape)
g_features = self.select_random_features(
gindividualf[j], args.rl_seq_len)
if not args.use_cpu:
q_features = q_features.cuda()
g_features = g_features.cuda()
# print(q_features.shape, g_features.shape)
self.env = Environment({
'features': q_features,
'id': qid
}, {
'features': g_features,
'id': gid
}, args.rp)
_, reward, iters, q_vals = self.play_one_episode(is_test=True)
instance_rewards[i, j] = reward
comparisons[i, j] = iters
distmat[i, j] = (q_vals[:, 1] - q_vals[:, 0]).item()
# break
print("Computing CMC and mAP (+ve distmat)")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
print("Computing CMC and mAP (-ve distmat)")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
print('average rewards', instance_rewards.mean().item())
print('average comparison', comparisons.mean().item())
return cmc[0]
def train():
model = FaceNetModel(embedding_size=args.embedding_size,
num_classes=args.num_classes).to(device)
optimizer = optim.Adam(model.parameters(), lr=args.learning_rate)
scheduler = lr_scheduler.StepLR(optimizer, step_size=50, gamma=0.1)
if args.start_epoch != 0:
checkpoint = torch.load(
'./log/checkpoint_epoch{}.pth'.format(args.start_epoch - 1))
model.load_state_dict(checkpoint['state_dict'])
train_loss = np.zeros((args.num_epochs))
train_accuracy = np.zeros((args.num_epochs))
for epoch in range(args.start_epoch,
args.num_epochs + args.start_epoch):
print(80 * '-')
print('Epoch [{}/{}]'.format(
epoch, args.num_epochs + args.start_epoch - 1))
data_transforms = {
'train':
transforms.Compose([
transforms.ToPILImage(),
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
]),
'valid':
transforms.Compose([
transforms.ToPILImage(),
transforms.ToTensor(),
transforms.Normalize(mean=[0.5, 0.5, 0.5],
std=[0.5, 0.5, 0.5])
])
}
face_dataset = {
'train':
TripletFaceDataset(root_dir=args.train_root_dir,
csv_name=args.train_csv_name,
num_triplets=args.num_train_triplets,
transform=data_transforms['train']),
'valid':
TripletFaceDataset(root_dir=args.valid_root_dir,
csv_name=args.valid_csv_name,
num_triplets=args.num_valid_triplets,
transform=data_transforms['valid'])
}
dataloaders = {
x: torch.utils.data.DataLoader(face_dataset[x],
batch_size=args.batch_size,
shuffle=False,
num_workers=args.num_workers)
for x in ['train', 'valid']
}
data_size = {x: len(face_dataset[x]) for x in ['train', 'valid']}
for phase in ['train', 'valid']:
labels, distances = [], []
triplet_loss_sum = 0.0
if phase == 'train':
scheduler.step()
model.train()
else:
model.eval()
for batch_idx, batch_sample in enumerate(dataloaders[phase]):
anc_img = batch_sample['anc_img'].to(device)
pos_img = batch_sample['pos_img'].to(device)
neg_img = batch_sample['neg_img'].to(device)
# print(anc_img.shape)
pos_cls = batch_sample['pos_class'].to(device)
neg_cls = batch_sample['neg_class'].to(device)
with torch.set_grad_enabled(phase == 'train'):
# anc_embed, pos_embed and neg_embed are embedding of image
anc_embed, pos_embed, neg_embed = model(
anc_img), model(pos_img), model(neg_img)
# print(anc_embed.shape)
# choose the hard negatives only for "training"
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist <
args.margin).cpu().numpy().flatten()
if phase == 'train':
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
else:
hard_triplets = np.where(all >= 0)
anc_hard_embed = anc_embed[hard_triplets].to(device)
pos_hard_embed = pos_embed[hard_triplets].to(device)
neg_hard_embed = neg_embed[hard_triplets].to(device)
anc_hard_img = anc_img[hard_triplets].to(device)
pos_hard_img = pos_img[hard_triplets].to(device)
neg_hard_img = neg_img[hard_triplets].to(device)
pos_hard_cls = pos_cls[hard_triplets].to(device)
neg_hard_cls = neg_cls[hard_triplets].to(device)
anc_img_pred = model.forward_classifier(
anc_hard_img).to(device)
pos_img_pred = model.forward_classifier(
pos_hard_img).to(device)
neg_img_pred = model.forward_classifier(
neg_hard_img).to(device)
triplet_loss = TripletLoss(args.margin).forward(
anc_hard_embed, pos_hard_embed,
neg_hard_embed).to(device)
if phase == 'train':
optimizer.zero_grad()
triplet_loss.backward()
optimizer.step()
dists = l2_dist.forward(anc_embed, pos_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
dists = l2_dist.forward(anc_embed, neg_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
triplet_loss_sum += triplet_loss.item()
torch.cuda.empty_cache()
avg_triplet_loss = triplet_loss_sum / data_size[phase]
labels = np.array(
[sublabel for label in labels for sublabel in label])
distances = np.array(
[subdist for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(
distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format(
phase, avg_triplet_loss))
print(' {} set - Accuracy = {:.8f}'.format(
phase, np.mean(accuracy)))
with open('./log/{}_log.txt'.format(phase), 'a') as f:
f.write(
str(epoch) + '\t' + str(np.mean(accuracy)) + '\t' +
str(avg_triplet_loss))
f.write("\n")
if phase == 'train':
torch.save(
{
'epoch': epoch,
'state_dict': model.state_dict()
}, 'log/checkpoint_epoch{}.pth'.format(epoch))
train_loss[epoch] = avg_triplet_loss
if phase == 'valid':
train_accuracy[epoch] = np.mean(accuracy)
print(80 * '-')
torch.save(model, 'model.pkl')
return train_loss, train_accuracy
def train(train_loader, model, optimizer, epoch):
# switch to train mode
model.train()
labels, distances = [], []
pbar = tqdm(enumerate(train_loader))
for batch_idx, (data_a, data_p, data_n, label_p, label_n) in pbar:
#print("on training{}".format(epoch))
if args.cuda:
data_a, data_p, data_n = data_a.cuda(), data_p.cuda(), data_n.cuda()
# compute output
out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)
if epoch > args.min_softmax_epoch:
triplet_loss = TripletMarginLoss(args.margin).forward(out_a, out_p, out_n)
loss = triplet_loss
# compute gradient and update weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
logger.log_value('selected_triplet_loss', triplet_loss.data.item()).step()
#logger.log_value('selected_cross_entropy_loss', cross_entropy_loss.data.item()).step()
logger.log_value('selected_total_loss', loss.data.item()).step()
if batch_idx % args.log_interval == 0:
pbar.set_description(
'Train Epoch: {:3d} [{:8d}/{:8d} ({:3.0f}%)]\tLoss: {:.6f}'.format(
# epoch, batch_idx * len(data_a), len(train_loader.dataset),
epoch, batch_idx * len(data_a), len(train_loader) * len(data_a),
100. * batch_idx / len(train_loader),
loss.data.item()))
dists = distance(out_a, out_n, args.distance)
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
dists = distance(out_a, out_p, args.distance)
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
else:
# Choose the hard negatives
d_p = distance(out_a, out_p, args.distance)
d_n = distance(out_a, out_n, args.distance)
all = (d_n - d_p < args.margin).cpu().data.numpy().flatten()
# log loss value for mini batch.
total_coorect = np.where(all == 0)
logger.log_value('Minibatch Train Accuracy', len(total_coorect[0]))
total_dist = (d_n - d_p).cpu().data.numpy().flatten()
logger.log_value('Minibatch Train distance', np.mean(total_dist))
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
if args.cuda:
out_selected_a = Variable(torch.from_numpy(out_a.cpu().data.numpy()[hard_triplets]).cuda())
out_selected_p = Variable(torch.from_numpy(out_p.cpu().data.numpy()[hard_triplets]).cuda())
out_selected_n = Variable(torch.from_numpy(out_n.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_a = Variable(torch.from_numpy(data_a.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_p = Variable(torch.from_numpy(data_p.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_n = Variable(torch.from_numpy(data_n.cpu().data.numpy()[hard_triplets]).cuda())
else:
out_selected_a = Variable(torch.from_numpy(out_a.data.numpy()[hard_triplets]))
out_selected_p = Variable(torch.from_numpy(out_p.data.numpy()[hard_triplets]))
out_selected_n = Variable(torch.from_numpy(out_n.data.numpy()[hard_triplets]))
selected_data_a = Variable(torch.from_numpy(data_a.data.numpy()[hard_triplets]))
selected_data_p = Variable(torch.from_numpy(data_p.data.numpy()[hard_triplets]))
selected_data_n = Variable(torch.from_numpy(data_n.data.numpy()[hard_triplets]))
selected_label_p = torch.from_numpy(label_p.cpu().numpy()[hard_triplets])
selected_label_n= torch.from_numpy(label_n.cpu().numpy()[hard_triplets])
triplet_loss = TripletMarginLoss(args.margin).forward(out_selected_a, out_selected_p, out_selected_n)
cls_a = model.forward_classifier(selected_data_a)
cls_p = model.forward_classifier(selected_data_p)
cls_n = model.forward_classifier(selected_data_n)
criterion = nn.CrossEntropyLoss()
predicted_labels = torch.cat([cls_a,cls_p,cls_n])
if args.cuda:
true_labels = torch.cat([Variable(selected_label_p.cuda()),Variable(selected_label_p.cuda()),Variable(selected_label_n.cuda())])
cross_entropy_loss = criterion(predicted_labels.cuda(),true_labels.cuda())
else:
true_labels = torch.cat([Variable(selected_label_p),Variable(selected_label_p),Variable(selected_label_n)])
cross_entropy_loss = criterion(predicted_labels,true_labels)
loss = cross_entropy_loss + triplet_loss * args.loss_ratio
# compute gradient and update weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
# log loss value for hard selected sample
logger.log_value('selected_triplet_loss', triplet_loss.data).step()
logger.log_value('selected_cross_entropy_loss', cross_entropy_loss.data).step()
logger.log_value('selected_total_loss', loss.data).step()
if batch_idx % args.log_interval == 0:
pbar.set_description(
'Train Epoch: {:3d} [{:8d}/{:8d} ({:3.0f}%)]\tLoss: {:.6f} \t Number of Selected Triplets: {:4d}'.format(
# epoch, batch_idx * len(data_a), len(train_loader.dataset),
epoch, batch_idx * len(data_a), len(train_loader) * len(data_a),
100. * batch_idx / len(train_loader),
loss.data,len(hard_triplets[0])))
dists = distance(out_selected_a, out_selected_n, args.distance)
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
dists = distance(out_selected_a, out_selected_p, args.distance)
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
#accuracy for hard selected sample, not all sample.
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array([subdist for dist in distances for subdist in dist])
tpr, fpr, accuracy = evaluate(distances, labels)
print('\33[91mTrain set: Accuracy: {:.8f}\n\33[0m'.format(np.mean(accuracy)))
logger.log_value('Train Accuracy', np.mean(accuracy))
# do checkpointing
torch.save({'epoch': epoch + 1, 'state_dict': model.state_dict(),
'optimizer': optimizer.state_dict()},
'{}/checkpoint_{}.pth'.format(LOG_DIR, epoch))
def train_valid(model, optimizer, scheduler, epoch, dataloaders, data_size,
start_time):
for phase in ['train', 'valid']:
if args.pure_validation and phase == 'train':
continue
if args.pure_training and phase == 'valid':
continue
labels, distances = [], []
triplet_loss_sum = 0.0
if phase == 'train':
scheduler.step()
model.train()
else:
model.eval()
#for batch_idx in range(0, data_size[phase], 1):
for batch_idx, batch_sample in enumerate(dataloaders[phase]):
#print("batch_idx:", batch_idx)
try:
#batch_sample = dataloaders[phase][batch_idx]
if not 'exception' in batch_sample:
anc_img = batch_sample['anc_img'].to(device)
pos_img = batch_sample['pos_img'].to(device)
neg_img = batch_sample['neg_img'].to(device)
pos_cls = batch_sample['pos_class'].to(device)
neg_cls = batch_sample['neg_class'].to(device)
with torch.set_grad_enabled(phase == 'train'):
# anc_embed, pos_embed and neg_embed are encoding(embedding) of image
anc_embed, pos_embed, neg_embed = model(
anc_img), model(pos_img), model(neg_img)
#for i in anc_embed:
# print(i.item())
if args.num_valid_triplets <= 100:
anc_embed_cpu = anc_embed.cpu()
pos_embed_cpu = pos_embed.cpu()
neg_embed_cpu = neg_embed.cpu()
pos_cls_cpu = pos_cls.cpu()
neg_cls_cpu = neg_cls.cpu()
pd.DataFrame([t.numpy() for t in anc_embed_cpu
]).to_csv("./embeddings.csv",
mode='a',
header=None)
pd.DataFrame([t.numpy() for t in pos_embed_cpu
]).to_csv("./embeddings.csv",
mode='a',
header=None)
pd.DataFrame([t.numpy() for t in neg_embed_cpu
]).to_csv("./embeddings.csv",
mode='a',
header=None)
pd.DataFrame({
'type':
"anc",
'id':
batch_sample['anc_id'],
'class':
pos_cls_cpu,
'train_set':
args.train_csv_name.split('.')[0],
'val_set':
args.valid_csv_name.split('.')[0]
}).to_csv("./embeddings_info.csv",
mode='a',
header=None)
pd.DataFrame({
'type':
"pos",
'id':
batch_sample['pos_id'],
'class':
pos_cls_cpu,
'train_set':
args.train_csv_name.split('.')[0],
'val_set':
args.valid_csv_name.split('.')[0]
}).to_csv("./embeddings_info.csv",
mode='a',
header=None)
pd.DataFrame({
'type':
"neg",
'id':
batch_sample['neg_id'],
'class':
pos_cls_cpu,
'train_set':
args.train_csv_name.split('.')[0],
'val_set':
args.valid_csv_name.split('.')[0]
}).to_csv("./embeddings_info.csv",
mode='a',
header=None)
#print([t.size() for t in anc_embed])
# choose the hard negatives only for "training"
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist <
args.margin).cpu().numpy().flatten()
if phase == 'train':
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
else:
hard_triplets = np.where(all >= 0)
anc_hard_embed = anc_embed[hard_triplets].to(device)
pos_hard_embed = pos_embed[hard_triplets].to(device)
neg_hard_embed = neg_embed[hard_triplets].to(device)
anc_hard_img = anc_img[hard_triplets].to(device)
pos_hard_img = pos_img[hard_triplets].to(device)
neg_hard_img = neg_img[hard_triplets].to(device)
pos_hard_cls = pos_cls[hard_triplets].to(device)
neg_hard_cls = neg_cls[hard_triplets].to(device)
anc_img_pred = model.forward_classifier(
anc_hard_img).to(device)
pos_img_pred = model.forward_classifier(
pos_hard_img).to(device)
neg_img_pred = model.forward_classifier(
neg_hard_img).to(device)
triplet_loss = TripletLoss(args.margin).forward(
anc_hard_embed, pos_hard_embed,
neg_hard_embed).to(device)
if phase == 'train':
optimizer.zero_grad()
triplet_loss.backward()
optimizer.step()
dists = l2_dist.forward(anc_embed, pos_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
dists = l2_dist.forward(anc_embed, neg_embed)
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
triplet_loss_sum += triplet_loss.item()
except:
#traceback.print_exc()
print("traceback: ", traceback.format_exc())
print("something went wrong with batch_idx: ", batch_idx,
", batch_sample:", batch_sample, ", neg_img size: ",
batch_sample['neg_img'].shape, ", pos_img size: ",
batch_sample['pos_img'].shape, ", anc_img size: ",
batch_sample['anc_img'].shape)
avg_triplet_loss = triplet_loss_sum / data_size[phase]
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array(
[subdist for dist in distances for subdist in dist])
nrof_pairs = min(len(labels), len(distances))
if nrof_pairs >= 10:
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format(
phase, avg_triplet_loss))
print(' {} set - Accuracy = {:.8f}'.format(
phase, np.mean(accuracy)))
duration = time.time() - start_time
with open('{}/{}_log_epoch{}.txt'.format(log_dir, phase, epoch),
'w') as f:
f.write(
str(epoch) + '\t' + str(np.mean(accuracy)) + '\t' +
str(avg_triplet_loss) + '\t' + str(duration))
if phase == 'train':
torch.save({
'epoch': epoch,
'state_dict': model.state_dict()
}, '{}/checkpoint_epoch{}.pth'.format(save_dir, epoch))
else:
plot_roc(fpr,
tpr,
figure_name='{}/roc_valid_epoch_{}.png'.format(
log_dir, epoch))
def test(model, queryloader, galleryloader, pool, use_gpu, ranks=[1, 5, 10, 20]):
with torch.no_grad():
model.eval()
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in tqdm(enumerate(queryloader), total=len(queryloader)):
if use_gpu:
imgs = imgs.cuda()
# imgs = Variable(imgs, volatile=True)
# b=1, n=number of clips, s=16
b, n, s, c, h, w = imgs.size()
assert (b == 1)
imgs = imgs.view(b * n, s, c, h, w)
features = model(imgs)
features = features.view(n, -1)
features = torch.mean(features, 0)
features = features.data.cpu().numpy()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
if batch_idx % 20 == 0:
gc.collect()
qf = np.asarray(qf, dtype=np.float32)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
gc.collect()
print("Extracted features for query set, obtained {}-by-{} matrix".format(qf.shape[0], qf.shape[1]))
gf, g_pids, g_camids = [], [], []
for batch_idx, (imgs, pids, camids) in tqdm(enumerate(galleryloader), total=len(galleryloader)):
if use_gpu:
imgs = imgs.cuda()
# imgs = Variable(imgs, volatile=True)
b, n, s, c, h, w = imgs.size()
imgs = imgs.view(b * n, s, c, h, w)
assert (b == 1)
features = model(imgs)
features = features.view(n, -1)
if pool == 'avg':
features = torch.mean(features, 0)
else:
features, _ = torch.max(features, 0)
features = features.data.cpu().numpy()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
if batch_idx % 20 == 0:
gc.collect()
gf = np.asarray(gf, dtype=np.float32)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
gc.collect()
print("Extracted features for gallery set, obtained {}-by-{} matrix".format(gf.shape[0], gf.shape[1]))
print("Computing distance matrix")
m, n = qf.shape[0], gf.shape[0]
distmat = np.tile(np.sum(np.power(qf, 2), axis=1, keepdims=True), (1, n)) + \
np.tile(np.sum(np.power(gf, 2), axis=1, keepdims=True), (1, m)).T
distmat -= 2 * blas.sgemm(1, qf, gf.T)
# distmat = np.power(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
# torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
# distmat.addmm_(1, -2, qf, gf.t())
# distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def test(model,
queryloader,
galleryloader,
pool,
use_gpu,
ranks=[1, 5, 10, 20]):
model.eval()
qf, qf_d, q_pids, q_camids = [], [], [], []
for batch_idx, (imgs, imgs_depths, pids, camids) in enumerate(queryloader):
if use_gpu:
imgs = imgs.cuda()
imgs_depths = imgs_depths.cuda()
imgs = Variable(imgs, volatile=True)
imgs_depths = Variable(imgs_depths, volatile=True)
# b=1, n=number of clips, s=16
b, n, s, c, h, w = imgs.size()
bd, nd, sd, cd, hd, wd = imgs_depths.size()
assert (b == 1)
assert (bd == 1)
imgs = imgs.view(b * n, s, c, h, w)
imgs_depths = imgs_depths.view(bd * nd, sd, cd, hd, wd)
features = model(imgs)
features = features.view(n, -1)
features = torch.mean(features, 0)
features = features.data.cpu()
features_d = features_d.view(nd, -1)
features_d = torch.mean(features_d, 0)
features_d = features_d.data.cpu()
qf.append(features)
qf_d.append(features_d)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.stack(qf)
qf_d = torch.stack(qf_d)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".format(
qf.size(0), qf.size(1)))
gf, gf_d, g_pids, g_camids = [], [], [], []
for batch_idx, (imgs, imgs_depths, pids,
camids) in enumerate(galleryloader):
if use_gpu:
imgs = imgs.cuda()
imgs_depths = imgs_depths.cuda()
imgs = Variable(imgs, volatile=True)
imgs_depths = Variable(imgs_depths, volatile=True)
b, n, s, c, h, w = imgs.size()
bd, nd, sd, cd, hd, wd = imgs_depths.size()
imgs = imgs.view(b * n, s, c, h, w)
imgs_depths = imgs_depths.view(bd * nd, sd, cd, hd, wd)
assert (b == 1)
assert (bd == 1)
features = model(imgs)
features = features.view(n, -1)
features_d = features_d.view(nd, -1)
if pool == 'avg':
features = torch.mean(features, 0)
features_d = torch.mean(features_d, 0)
else:
features, _ = torch.max(features, 0)
features_d, _ = torch.max(features_d, 0)
features = features.data.cpu()
features_d = features_d.data.cpu()
gf.append(features)
gf_d.append(features_d)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.stack(gf)
gf_d = torch.stack(gf_d)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print(
"Extracted features for gallery set, obtained {}-by-{} matrix".format(
gf.size(0), gf.size(1)))
print("Computing distance matrix")
m, n, md, nd = qf.size(0), gf.size(0), qf_d.size(0), gf_d.size(
0) #rimettere depth
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
distmat = distmat.numpy()
distmat_d = torch.pow(qf_d, 2).sum(dim=1, keepdim=True).expand(md, nd) + \
torch.pow(gf_d, 2).sum(dim=1, keepdim=True).expand(nd, md).t()
distmat_d.addmm_(1, -2, qf_d, gf_d.t())
distmat_d = distmat_d.numpy()
distmat_tot = (distmat + distmat_d) / 2
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat_tot, q_pids, g_pids, q_camids,
g_camids) #rimettere distmat_tot
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def test(model,
classifier_model,
queryloader,
galleryloader,
pool,
use_gpu,
ranks=[1, 5, 10, 20]):
model.eval()
classifier_model.eval()
qf, q_pids, q_camids = [], [], []
tot_part = args.part1 + args.part2
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
if use_gpu:
imgs = imgs[:, :40, :, :, :, :]
imgs = imgs.cuda()
with torch.no_grad():
#imgs = Variable(imgs)
# b=1, n=number of clips, s=16
b, n, s, c, h, w = imgs.size()
assert (b == 1)
features = []
m = n
if n > 40:
m = 40
for i in range(m):
b, parts1, parts2 = model(imgs[:, i, :, :, :, :])
features.append(classifier_model(b, parts1, parts2, 1, s))
features = torch.cat(features, 0)
features = features.data.cpu()
features = features.view(m, 1024, tot_part + 1)
fnorm = torch.norm(features, p=2, dim=1,
keepdim=True) * np.sqrt(tot_part + 1)
features = features.div(fnorm.expand_as(features))
features = features.view(features.size(0), -1)
features = torch.mean(features, 0)
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.stack(qf)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".format(
qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu:
imgs = imgs[:, :80, :, :, :, :]
imgs = imgs.cuda()
with torch.no_grad():
imgs = Variable(imgs)
b, n, s, c, h, w = imgs.size()
features = []
for i in range(n):
b, parts1, parts2 = model(imgs[:, i, :, :, :, :])
features.append(classifier_model(b, parts1, parts2, 1, s))
features = torch.cat(features, 0)
features = features.view(n, 1024, tot_part + 1)
fnorm = torch.norm(features, p=2, dim=1,
keepdim=True) * np.sqrt(tot_part + 1)
features = features.div(fnorm.expand_as(features))
features = features.view(features.size(0), -1)
if pool == 'avg':
features = torch.mean(features, 0)
else:
features, _ = torch.max(features, 0)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
torch.cuda.empty_cache()
gf = torch.stack(gf)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print(
"Extracted features for gallery set, obtained {}-by-{} matrix".format(
gf.size(0), gf.size(1)))
print("Computing distance matrix")
m, n = qf.size(0), gf.size(0)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(qf, gf.t(), beta=1, alpha=-2)
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
# re-ranking
from person_re_ranking.python_version.re_ranking_feature import re_ranking
rerank_distmat = re_ranking(qf.numpy(),
gf.numpy(),
k1=20,
k2=6,
lambda_value=0.3)
print("Computing CMC and mAP for re-ranking")
cmc, mAP = evaluate(rerank_distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def test(model, queryloader, galleryloader, use_gpu, ranks=[1, 5, 10, 20]):
since = time.time()
model.eval()
qf, q_pids, q_camids = [], [], []
for batch_idx, (vids, pids, camids, img_paths) in enumerate(queryloader):
if use_gpu:
vids = vids.cuda()
feat = model(vids) #[b, c * n]
feat_list = torch.split(feat, [bn.num_features for bn in model.module.bn], dim=1)
norm_feat_list = []
for i, f in enumerate(feat_list):
f = model.module.bn[i](f) #[bs, c]
f = F.normalize(f, p=2, dim=1, eps=1e-12)
norm_feat_list.append(f)
feat = torch.cat(norm_feat_list, 1)
qf.append(feat)
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {} matrix".format(qf.shape))
gf, g_pids, g_camids = [], [], []
for batch_idx, (vids, pids, camids, img_paths) in enumerate(galleryloader):
if use_gpu:
vids = vids.cuda()
feat = model(vids)
feat_list = torch.split(feat, [bn.num_features for bn in model.module.bn], dim=1)
norm_feat_list = []
for i, f in enumerate(feat_list):
f = model.module.bn[i](f) #[bs, c]
f = F.normalize(f, p=2, dim=1, eps=1e-12)
norm_feat_list.append(f)
feat = torch.cat(norm_feat_list, 1)
gf.append(feat)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
if args.dataset == 'mars':
# gallery set must contain query set, otherwise 140 query imgs will not have ground truth.
gf = torch.cat((qf, gf), 0)
g_pids = np.append(q_pids, g_pids)
g_camids = np.append(q_camids, g_camids)
print("Extracted features for gallery set, obtained {} matrix".format(gf.shape))
time_elapsed = time.time() - since
print('Extracting features complete in {:.0f}m {:.0f}s'.format(time_elapsed // 60, time_elapsed % 60))
print("Computing distance matrix")
distmat = - torch.mm(qf, gf.t())
distmat = distmat.data.cpu()
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP, rank5s = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.2%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.2%}".format(r, cmc[r-1]))
print("------------------")
number_wrong_rank1 = 0
for q_idx in range(len(q_pids)):
if g_pids[rank5s[q_idx][0]] == q_pids[q_idx]:
continue # pass for correct answers
#print("Wrong rank1 query {}; gallery {}".format(q_pids[q_idx], g_pids[rank5s[q_idx]]))
number_wrong_rank1 += 1
print("#Wrong %rank1 {}".format(number_wrong_rank1 * 1.0 / len(rank5s)))
return cmc[0]
def test(model, queryloader, galleryloader, use_gpu, ranks=[1, 5, 10, 20]):
model.eval()
qf, q_pids, q_camids = [], [], []
print('extracting query feats')
for batch_idx, (imgs, pids, camids) in enumerate(tqdm(queryloader)):
if use_gpu:
imgs = imgs.cuda()
with torch.no_grad():
#imgs = Variable(imgs, volatile=True)
# b=1, n=number of clips, s=16
b, n, s, c, h, w = imgs.size()
assert(b == 1)
imgs = imgs.view(b*n, s, c, h, w)
features = get_features(model, imgs, args.test_num_tracks)
features = torch.mean(features, 0)
features = features.data.cpu()
qf.append(features)
q_pids.extend(pids)
q_camids.extend(camids)
torch.cuda.empty_cache()
qf = torch.stack(qf)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
print('extracting gallery feats')
for batch_idx, (imgs, pids, camids) in enumerate(tqdm(galleryloader)):
if use_gpu:
imgs = imgs.cuda()
with torch.no_grad():
#imgs = Variable(imgs, volatile=True)
b, n, s, c, h, w = imgs.size()
imgs = imgs.view(b*n, s, c, h, w)
assert(b == 1)
# handle chunked data
features = get_features(model, imgs, args.test_num_tracks)
features = torch.mean(features, 0)
torch.cuda.empty_cache()
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.stack(gf)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print("Extracted features for gallery set, obtained {}-by-{} matrix".format(gf.size(0), gf.size(1)))
print("Computing distance matrix")
m, n = qf.size(0), gf.size(0)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
distmat.addmm_(1, -2, qf, gf.t())
distmat = distmat.numpy()
print("Computing CMC and mAP")
cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r-1]))
print("------------------")
# re-ranking
from person_re_ranking.python_version.re_ranking_feature import re_ranking
rerank_distmat = re_ranking(
qf.numpy(), gf.numpy(), k1=20, k2=6, lambda_value=0.3)
print("Computing CMC and mAP for re-ranking")
cmc, mAP = evaluate(rerank_distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r-1]))
print("------------------")
return cmc[0], mAP
def train(train_loader, model, optimizer, epoch):
# switch to train mode
model.train()
pbar = tqdm(enumerate(train_loader))
labels, distances = [], []
for batch_idx, (data_a, data_p, data_n, label_p, label_n) in pbar:
data_a, data_p, data_n = data_a.cuda(), data_p.cuda(), data_n.cuda()
data_a, data_p, data_n = Variable(data_a), Variable(data_p), \
Variable(data_n)
# compute output
out_a, out_p, out_n = model(data_a), model(data_p), model(data_n)
# Choose the hard negatives
d_p = l2_dist.forward(out_a, out_p)
d_n = l2_dist.forward(out_a, out_n)
all = (d_n - d_p < args.margin).cpu().data.numpy().flatten()
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
out_selected_a = Variable(torch.from_numpy(out_a.cpu().data.numpy()[hard_triplets]).cuda())
out_selected_p = Variable(torch.from_numpy(out_p.cpu().data.numpy()[hard_triplets]).cuda())
out_selected_n = Variable(torch.from_numpy(out_n.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_a = Variable(torch.from_numpy(data_a.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_p = Variable(torch.from_numpy(data_p.cpu().data.numpy()[hard_triplets]).cuda())
selected_data_n = Variable(torch.from_numpy(data_n.cpu().data.numpy()[hard_triplets]).cuda())
selected_label_p = torch.from_numpy(label_p.cpu().numpy()[hard_triplets])
selected_label_n = torch.from_numpy(label_n.cpu().numpy()[hard_triplets])
triplet_loss = TripletMarginLoss(args.margin).forward(out_selected_a, out_selected_p, out_selected_n)
cls_a = model.forward_classifier(selected_data_a)
cls_p = model.forward_classifier(selected_data_p)
cls_n = model.forward_classifier(selected_data_n)
criterion = nn.CrossEntropyLoss()
predicted_labels = torch.cat([cls_a, cls_p, cls_n])
true_labels = torch.cat(
[Variable(selected_label_p.cuda()), Variable(selected_label_p.cuda()), Variable(selected_label_n.cuda())])
cross_entropy_loss = criterion(predicted_labels.cuda(), true_labels.cuda())
loss = cross_entropy_loss + triplet_loss
# compute gradient and update weights
optimizer.zero_grad()
loss.backward()
optimizer.step()
# update the optimizer learning rate
adjust_learning_rate(optimizer)
# log loss value
logger.log_value('triplet_loss', triplet_loss.data[0]).step()
logger.log_value('cross_entropy_loss', cross_entropy_loss.data[0]).step()
logger.log_value('total_loss', loss.data[0]).step()
if batch_idx % args.log_interval == 0:
pbar.set_description(
'Train Epoch: {} [{}/{} ({:.0f}%)]\tLoss: {:.6f} \t # of Selected Triplets: {}'.format(
epoch, batch_idx * len(data_a), len(train_loader.dataset),
100. * batch_idx / len(train_loader),
loss.data[0], len(hard_triplets[0])))
dists = l2_dist.forward(out_selected_a,
out_selected_n) # torch.sqrt(torch.sum((out_a - out_n) ** 2, 1)) # euclidean distance
distances.append(dists.data.cpu().numpy())
labels.append(np.zeros(dists.size(0)))
dists = l2_dist.forward(out_selected_a,
out_selected_p) # torch.sqrt(torch.sum((out_a - out_p) ** 2, 1)) # euclidean distance
distances.append(dists.data.cpu().numpy())
labels.append(np.ones(dists.size(0)))
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array([subdist[0] for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print('\33[91mTrain set: Accuracy: {:.8f}\n\33[0m'.format(np.mean(accuracy)))
logger.log_value('Train Accuracy', np.mean(accuracy))
plot_roc(fpr, tpr, figure_name="roc_train_epoch_{}.png".format(epoch))
# do checkpointing
torch.save({'epoch': epoch + 1, 'state_dict': model.state_dict()},
'{}/checkpoint_{}.pth'.format(LOG_DIR, epoch))
def test(model, queryloader, galleryloader, use_gpu, ranks=[1, 5, 10, 20]):
batch_time = AverageMeter()
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
if use_gpu: imgs = imgs.cuda()
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
# print('features', features.size(), features)
qf.append(features)
# print('qf', len(qf))
q_pids.extend(pids)
q_camids.extend(camids)
qf = torch.cat(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".
format(qf.size(0), qf.size(1)))
gf, g_pids, g_camids = [], [], []
end = time.time()
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu: imgs = imgs.cuda()
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.data.cpu()
gf.append(features)
g_pids.extend(pids)
g_camids.extend(camids)
gf = torch.cat(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print("Extracted features for gallery set, obtained {}-by-{} matrix".
format(gf.size(0), gf.size(1)))
print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(
batch_time.avg, args.test_batch))
# print('qf', qf.size(), qf)
# print('gf', gf.size(), gf)
m, n = qf.size(0), gf.size(0)
# print('m,n',m,n)
# print('qf', qf.size(), qf)
# qf_pow = torch.pow(qf, 2)
# # print('qf_pow', qf_pow.size(), qf_pow)
# qf_sum = qf_pow.sum(dim=1, keepdim=True)
# # print('qf_sum', qf_sum.size(), qf_sum)
# qf_exp = qf_sum.expand(m, n)
# # print('qf_exp', qf_exp.size(), qf_exp)
# # print('gf', gf.size(), gf)
# gf_pow = torch.pow(gf, 2)
# # print('gf_pow', gf_pow.size(), gf_pow)
# gf_sum = gf_pow.sum(dim=1, keepdim=True)
# # print('gf_sum', gf_sum.size(), gf_sum)
# gf_exp = gf_sum.expand(n, m)
# # print('gf_exp', gf_exp.size(), gf_exp)
# gf_t = gf_exp.t()
# print('gf_t', gf_t.size(), gf_t)
distmat = torch.pow(qf, 2).sum(dim=1, keepdim=True).expand(m, n) + \
torch.pow(gf, 2).sum(dim=1, keepdim=True).expand(n, m).t()
# distmat = qf_exp + gf_t
# # print('distmat', distmat.size(), distmat)
# # print('mm', distmat.size, qf.size, gf.size, gf.t().size)
# qf = qf.numpy()
# gf = gf.numpy()
# mm = np.dot(qf, gf.T)
# # mm = torch.mm(qf, gf.t())
# print('mm', mm.shape, distmat.shape, qf.shape, gf.T.shape)
# distmat = distmat.numpy() + mm*(-2)
distmat.addmm_(1, -2, qf, gf.t())
# print('distmat', distmat.shape, distmat)
distmat = distmat.numpy()
print("Computing CMC and mAP")
# cmc, mAP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids, use_metric_cuhk03=args.use_metric_cuhk03)
cmc, mAP = evaluate(distmat,
q_pids,
g_pids,
q_camids,
g_camids,
dataset_type=args.dataset)
print("Results ----------")
print("mAP: {:.1%}".format(mAP))
print("CMC curve")
for r in ranks:
print("Rank-{:<3}: {:.1%}".format(r, cmc[r - 1]))
print("------------------")
return cmc[0]
def test(model,
queryloader,
galleryloader,
pool,
use_gpu,
ranks=[1, 5, 10, 20],
return_distmat=False):
batch_time = AverageMeter()
model.eval()
with torch.no_grad():
qf, q_pids, q_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(queryloader):
if use_gpu: imgs = imgs.cuda()
b, n, s, c, h, w = imgs.size()
assert (b == 1)
imgs = imgs.view(b * n, s, c, h, w)
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.view(n, -1)
features = torch.mean(features, 0)
features = features.cpu()
qf.append(features.numpy())
q_pids.extend(pids.numpy())
q_camids.extend(camids.numpy())
qf = np.stack(qf, 0)
q_pids = np.asarray(q_pids)
q_camids = np.asarray(q_camids)
print("Extracted features for query set, obtained {}-by-{} matrix".
format(qf.shape[0], qf.shape[1]))
gf, g_pids, g_camids = [], [], []
for batch_idx, (imgs, pids, camids) in enumerate(galleryloader):
if use_gpu: imgs = imgs.cuda()
b, n, s, c, h, w = imgs.size()
assert (b == 1)
imgs = imgs.view(b * n, s, c, h, w)
end = time.time()
features = model(imgs)
batch_time.update(time.time() - end)
features = features.view(n, -1)
features = torch.mean(features, 0)
features = features.cpu()
gf.append(features.numpy())
g_pids.extend(pids.numpy())
g_camids.extend(camids.numpy())
gf = np.stack(gf, 0)
g_pids = np.asarray(g_pids)
g_camids = np.asarray(g_camids)
print("Extracted features for gallery set, obtained {}-by-{} matrix".
format(gf.shape[0], gf.shape[1]))
print("==> BatchTime(s)/BatchSize(img): {:.3f}/{}".format(
batch_time.avg, args.test_batch * args.seq_len))
m, n = qf.shape[0], gf.shape[0]
distmat = np.tile((qf**2).sum(axis=1, keepdims=True),(1, n)) + \
np.tile((gf**2).sum(axis=1, keepdims=True),(1, m)).transpose()
distmat = distmat - 2 * np.dot(qf, gf.transpose())
print("Computing CMC and mAP")
cmc, mAP, mINP = evaluate(distmat, q_pids, g_pids, q_camids, g_camids)
print("Results ----------")
print("mINP: {:.1%} mAP: {:.1%} CMC curve {:.1%} {:.1%} {:.1%} {:.1%}".
format(mINP, mAP, cmc[1 - 1], cmc[5 - 1], cmc[10 - 1], cmc[20 - 1]))
print("------------------")
if return_distmat:
return distmat
return cmc[0]
def train_valid(model, optimizer, triploss, scheduler, epoch, dataloaders, data_size):
for phase in ['train', 'valid']:
labels, distances = [], []
triplet_loss_sum = 0.0
if phase == 'train':
scheduler.step()
if scheduler.last_epoch % scheduler.step_size == 0:
print("LR decayed to:", ', '.join(map(str, scheduler.get_lr())))
model.train()
else:
model.eval()
for batch_idx, batch_sample in enumerate(dataloaders[phase]):
anc_img = batch_sample['anc_img'].to(device)
pos_img = batch_sample['pos_img'].to(device)
neg_img = batch_sample['neg_img'].to(device)
# pos_cls = batch_sample['pos_class'].to(device)
# neg_cls = batch_sample['neg_class'].to(device)
with torch.set_grad_enabled(phase == 'train'):
# anc_embed, pos_embed and neg_embed are encoding(embedding) of image
anc_embed, pos_embed, neg_embed = model(anc_img), model(pos_img), model(neg_img)
# choose the semi hard negatives only for "training"
pos_dist = l2_dist.forward(anc_embed, pos_embed)
neg_dist = l2_dist.forward(anc_embed, neg_embed)
all = (neg_dist - pos_dist < args.margin).cpu().numpy().flatten()
if phase == 'train':
hard_triplets = np.where(all == 1)
if len(hard_triplets[0]) == 0:
continue
else:
hard_triplets = np.where(all >= 0)
anc_hard_embed = anc_embed[hard_triplets]
pos_hard_embed = pos_embed[hard_triplets]
neg_hard_embed = neg_embed[hard_triplets]
anc_hard_img = anc_img[hard_triplets]
pos_hard_img = pos_img[hard_triplets]
neg_hard_img = neg_img[hard_triplets]
# pos_hard_cls = pos_cls[hard_triplets]
# neg_hard_cls = neg_cls[hard_triplets]
model.module.forward_classifier(anc_hard_img)
model.module.forward_classifier(pos_hard_img)
model.module.forward_classifier(neg_hard_img)
triplet_loss = triploss.forward(anc_hard_embed, pos_hard_embed, neg_hard_embed)
if phase == 'train':
optimizer.zero_grad()
triplet_loss.backward()
optimizer.step()
distances.append(pos_dist.data.cpu().numpy())
labels.append(np.ones(pos_dist.size(0)))
distances.append(neg_dist.data.cpu().numpy())
labels.append(np.zeros(neg_dist.size(0)))
triplet_loss_sum += triplet_loss.item()
avg_triplet_loss = triplet_loss_sum / data_size[phase]
labels = np.array([sublabel for label in labels for sublabel in label])
distances = np.array([subdist for dist in distances for subdist in dist])
tpr, fpr, accuracy, val, val_std, far = evaluate(distances, labels)
print(' {} set - Triplet Loss = {:.8f}'.format(phase, avg_triplet_loss))
print(' {} set - Accuracy = {:.8f}'.format(phase, np.mean(accuracy)))
time = datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")
lr = '_'.join(map(str, scheduler.get_lr()))
layers = '+'.join(args.unfreeze.split(','))
write_csv(f'log/{phase}.csv', [time, epoch, np.mean(accuracy), avg_triplet_loss, layers, args.batch_size, lr])
if phase == 'valid':
save_last_checkpoint({'epoch': epoch,
'state_dict': model.module.state_dict(),
'optimizer_state': optimizer.state_dict(),
'accuracy': np.mean(accuracy),
'loss': avg_triplet_loss
})
save_if_best({'epoch': epoch,
'state_dict': model.module.state_dict(),
'optimizer_state': optimizer.state_dict(),
'accuracy': np.mean(accuracy),
'loss': avg_triplet_loss
}, np.mean(accuracy))
else:
plot_roc(fpr, tpr, figure_name='./log/roc_valid_epoch_{}.png'.format(epoch))