def visualize_ranked_images(config, base, loaders, dataset):
base.set_eval()
# init dataset
if dataset == 'market':
_datasets = [loaders.query_samples.samples, loaders.gallery_samples.samples]
_loaders = [loaders.query_loader, loaders.gallery_loader]
save_visualize_path = base.save_visualize_market_path
elif dataset == 'duke':
_datasets = [loaders.duke_query_samples.samples, loaders.duke_gallery_samples.samples]
_loaders = [loaders.duke_query_loader, loaders.duke_gallery_loader]
save_visualize_path = base.save_visualize_duke_path
# compute featuress
query_features, query_features2, gallery_features, gallery_features2 = compute_features(base, _loaders, True)
# compute cosine similarity
cosine_similarity = cosine_dist(
torch.tensor(query_features).cuda(),
torch.tensor(gallery_features).cuda()).data.cpu().numpy()
# compute verification score
ver_scores = compute_ver_scores(cosine_similarity, query_features2, gallery_features2, base.verificator, topk=25, sort='descend')
# visualize
visualize_ranked_results(cosine_similarity, ver_scores, _datasets, save_dir=save_visualize_path, topk=20, sort='descend')
def visualize(config, base, loaders):
base.set_eval()
# meters
query_features_meter, query_pids_meter, query_cids_meter = CatMeter(
), CatMeter(), CatMeter()
gallery_features_meter, gallery_pids_meter, gallery_cids_meter = CatMeter(
), CatMeter(), CatMeter()
# init dataset
if config.visualize_dataset == 'market':
_datasets = [
loaders.market_query_samples.samples,
loaders.market_gallery_samples.samples
]
_loaders = [loaders.market_query_loader, loaders.market_gallery_loader]
elif config.visualize_dataset == 'duke':
_datasets = [
loaders.duke_query_samples.samples,
loaders.duke_gallery_samples.samples
]
_loaders = [loaders.duke_query_loader, loaders.duke_gallery_loader]
elif config.visualize_dataset == 'customed':
_datasets = [loaders.query_samples, loaders.gallery_samples]
_loaders = [loaders.query_loader, loaders.gallery_loader]
# compute query and gallery features
with torch.no_grad():
for loader_id, loader in enumerate(_loaders):
for data in loader:
# compute feautres
images, pids, cids = data
features = base.model(images)
# save as query features
if loader_id == 0:
query_features_meter.update(features.data)
query_pids_meter.update(pids)
query_cids_meter.update(cids)
# save as gallery features
elif loader_id == 1:
gallery_features_meter.update(features.data)
gallery_pids_meter.update(pids)
gallery_cids_meter.update(cids)
# compute distance
query_features = query_features_meter.get_val()
gallery_features = gallery_features_meter.get_val()
distance = cosine_dist(query_features, gallery_features).data.cpu().numpy()
# visualize
visualize_ranked_results(distance,
_datasets,
config.visualize_output_path,
mode=config.visualize_mode)
def inference_buildgallery(config, base, loaders):
base.set_eval()
dir_q = '/home/teresa/reID/temp'
q_list = sorted(glob.glob(dir_q + '/*'))
for i, q in enumerate(q_list):
print('query', i, q)
loaders.set_query(q, 0, 1) #TODO: mod for multiple queries
loaders.process_query_data()
features_g = None
with torch.no_grad():
for batch_num, data in enumerate(loaders.query_loader):
images, pids, cids = data
features_q = base.model(images) #can be multiple features
if len(loaders.gallery) > 0:
for g in loaders.gallery:
path, feat, pid, cid = g
if features_g is None:
features_g = feat.unsqueeze(0)
else:
features_g = torch.cat([features_g, feat.unsqueeze(0)], dim=0)
#extract feat from tuples and convert to torch
#print(features_g)
distance = cosine_dist(features_q, features_g).data.cpu().numpy()
print(distance)
indices = np.argsort(distance, axis=1)[:,::-1] #smallest distance first, reverse order when add to gallery
print(indices)
closest_dist = distance[0][indices[0][0]]
print('closest match with dist',loaders.gallery[indices[0][0]][0], distance[0][indices[0][0]])
if closest_dist > 0.6:
pids = [loaders.gallery[indices[0][0]][2]] #2 for pid
else:
pids = [i]
#pids = assign_pid(distance,loaders.gallery) #one query has one row of distance hence
#another function to sort with index and assign PID
#for f in features_q: #TODO: mod for muliple queries
else:
pids = [i] #first pid for empty gallery
for fp in zip(features_q,pids):
loaders.add_gallery_all(q, fp[0], fp[1], 1)
print(loaders.gallery)
def mining_hard_pairs(feature_vector_list, pids):
'''
use global feature (the last one) to mining hard positive and negative pairs
cosine distance is used to measure similarity
:param feature_vector_list:
:param pids:
:return:
'''
global_feature_vectors = feature_vector_list[-1]
dist_matrix = cosine_dist(global_feature_vectors, global_feature_vectors)
label_matrix = label2similarity(pids, pids).float()
# 获得所有难得正样本的下标
_, sorted_mat_distance_index = torch.sort(dist_matrix + (9999999.) *
(1 - label_matrix),
dim=1,
descending=False)
hard_p_index = sorted_mat_distance_index[:, 0]
# 获得所有难得负样本的下标
_, sorted_mat_distance_index = torch.sort(dist_matrix + (-9999999.) *
(label_matrix),
dim=1,
descending=True)
hard_n_index = sorted_mat_distance_index[:, 0]
new_feature_vector_list = []
p_feature_vector_list = []
n_feature_vector_list = []
for feature_vector in feature_vector_list:
feature_vector = copy.copy(feature_vector)
new_feature_vector_list.append(feature_vector.detach())
feature_vector = copy.copy(feature_vector.detach())
p_feature_vector_list.append(feature_vector[hard_p_index, :])
feature_vector = copy.copy(feature_vector.detach())
n_feature_vector_list.append(feature_vector[hard_n_index, :])
return new_feature_vector_list, p_feature_vector_list, n_feature_vector_list
def visualize(config, base, loaders):
base.set_eval()
dir_q = '/home/teresa/reID/output_filter_2019-08-04-15-45-ip-101.flv/'
q_list = glob.glob(dir_q + '/*')
for i, q in enumerate(q_list):
loaders.add_query(q, 0, 1)
print(i, q)
# if i>100:
# break
dir_g = '/home/teresa/reID/output_filter_2019-08-04-15-45-ip-102.flv/'
g_list = glob.glob(dir_g + '/*')
for i, g in enumerate(g_list):
loaders.add_gallery(g, 0, 1)
print(i, g)
dir_g = '/home/teresa/reID/output_filter_2019-08-04-15-45-ip-103.flv/'
g_list = glob.glob(dir_g + '/*')
for i, g in enumerate(g_list):
loaders.add_gallery(g, 0, 1)
print(i, g)
dir_g = '/home/teresa/reID/output_filter_2019-08-04-15-45-ip-104.flv/'
g_list = glob.glob(dir_g + '/*')
for i, g in enumerate(g_list):
loaders.add_gallery(g, 0, 1)
print(i, g)
#call this when done adding
loaders.process_data()
# meters
query_features_meter, query_pids_meter, query_cids_meter = CatMeter(), CatMeter(), CatMeter()
gallery_features_meter, gallery_pids_meter, gallery_cids_meter = CatMeter(), CatMeter(), CatMeter()
# init dataset
if config.visualize_dataset == 'market':
_datasets = [loaders.market_query_samples.samples, loaders.market_gallery_samples.samples]
_loaders = [loaders.market_query_loader, loaders.market_gallery_loader]
elif config.visualize_dataset == 'duke':
_datasets = [loaders.duke_query_samples.samples, loaders.duke_gallery_samples.samples]
_loaders = [loaders.duke_query_loader, loaders.duke_gallery_loader]
elif config.visualize_dataset == 'customed':
_datasets = [loaders.query_samples, loaders.gallery_samples]
_loaders = [loaders.query_loader, loaders.gallery_loader]
print('query and then gallery',_datasets)
# compute query and gallery features
with torch.no_grad():
for loader_id, loader in enumerate(_loaders):
for batch_num, data in enumerate(loader):
# compute feautres
images, pids, cids = data
print('image len',len(images))
#print('loading all query followed by gallery images',len(pids), batch_num)
start_time = time.time()
features = base.model(images)
print('1 time get features',time.time()-start_time)
start_time = time.time()
features = base.model(images)
print('2 time get features',time.time()-start_time)
start_time = time.time()
features = base.model(images)
print('3 time get features',time.time()-start_time)
start_time = time.time()
features = base.model(images)
print('4 time get features',time.time()-start_time)
start_time = time.time()
features = base.model(images)
print('5 time get features',time.time()-start_time)
start_time = time.time()
features = base.model(images)
print('6 time get features',time.time()-start_time)
# save as query features
if loader_id == 0:
query_features_meter.update(features.data)
query_pids_meter.update(pids)
query_cids_meter.update(cids)
# save as gallery features
elif loader_id == 1:
gallery_features_meter.update(features.data)
gallery_pids_meter.update(pids)
gallery_cids_meter.update(cids)
img_path = ''
# compute distance
query_features = query_features_meter.get_val()
gallery_features = gallery_features_meter.get_val()
start_time = time.time()
distance = cosine_dist(query_features, gallery_features).data.cpu().numpy()
print('time get dist',time.time()-start_time)
print(len(distance),len(distance),type(distance))
print('dist size',len(distance),len(distance[0]))
# visualize
visualize_ranked_results(distance, _datasets, config.visualize_output_path, mode=config.visualize_mode, topk=20)