def inference(self,conf,img):
mtcnn = MTCNN()
learner = face_learner(conf,True)
learner.load_state(conf,'final.pth',True,True)
learner.model.eval()
targets, names = load_facebank(conf)
image = Image.open(img)
frame = cv2.imread(img,cv2.IMREAD_COLOR)
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
try:
bboxes, faces = mtcnn.align_multi(image, conf.face_limit, conf.min_face_size)
bboxes = bboxes[:,:-1] #shape:[10,4],only keep 10 highest possibiity faces
bboxes = bboxes.astype(int)
bboxes = bboxes + [-1,-1,1,1] # personal choice
results, score = learner.infer(conf, faces, targets, False)
name = names[results[0]+1]
frame = draw_box_name(bboxes[0], name, frame)
except Exception as ex:
name = "Can't detect face."
h, w, c = frame.shape
bbox = [int(h*0.5),int(w*0.5),int(h*0.5),int(w*0.5)]
frame = draw_box_name(bbox, name, frame)
return name, frame
def authenticuser(path,userid):
conf = get_config(False)
mtcnn = MTCNN()
print('mtcnn loaded')
learner = face_learner(conf, True)
learner.threshold = 1.35
learner.load_state(conf, 'cpu_final.pth', True, True)
learner.model.eval()
print('learner loaded')
targets = load_facebank_user(conf,userid)
names=['Unknown',userid]
print('facebank loaded')
count =0
while True:
frame = cv2.imread(path)
#try:
image = Image.fromarray(frame)
bboxes, faces = mtcnn.align_multi(image, conf.face_limit, conf.min_face_size)
bboxes = bboxes[:,:-1] #shape:[10,4],only keep 10 highest possibiity faces
bboxes = bboxes.astype(int)
bboxes = bboxes + [-1,-1,1,1] # personal choice
results, score = learner.infer(conf, faces, targets)
for idx,bbox in enumerate(bboxes):
frame = draw_box_name(bbox, names[results[idx] + 1] + '_{:.2f}'.format(100-score[idx]), frame)
result={"_result":"success", "User Verified with":{"confidence": '{:.2f}%'.format(100-score[idx]), "userid": names[results[idx] + 1] , "error": "Success"}}
accuracy.append('{:.2f}'.format(100-score[idx]))
user.append(names[results[idx] + 1])
print( names[results[idx] + 1],'{:.2f}'.format(100-score[idx]))
count=1
#except:
# print('detect error')
if count>0:
break
return result
def load_Learner(conf, args):
learner = face_learner(conf, True)
learner.threshold = args.threshold
if conf.device.type == 'cpu':
learner.load_state(conf, 'cpu_final.pth', True, True)
else:
learner.load_state(conf, 'final.pth', True, True)
learner.model.eval()
return learner
def __init__(self):
self.faceDataDir = './data/facerec/'
self.modelDir = './data/desired_model/'
self.modelFn = self.modelDir + 'model_mobilefacenet.pth'
self.learner = face_learner(use_mobilfacenet=True)
self.learner.load_state(self.modelFn)
self.learner.model.eval()
self.allPersons = []
self.read_all_register_info()
def train():
conf = get_config(False)
mtcnn = MTCNN()
#print('mtcnn loaded')
learner = face_learner(conf, True)
learner.load_state(conf, 'cpu_final.pth', True, True)
learner.model.eval()
print('learner loaded')
targets, names = prepare_facebank(conf, learner.model, mtcnn)
return {'_result': 'success', '_message': 'Model Is Updated'}
def extract2db(dbname):
import tqdm
ds = DatasetIJBC()
loader = torch.utils.data.DataLoader(ds,
batch_size=1,
num_workers=12,
shuffle=False,
pin_memory=True,
collate_fn=my_collate)
learner = face_learner(conf, inference=True)
assert conf.device.type != 'cpu'
prefix = list(model_path.glob('model*_*.pth'))[0].name.replace(
'model_', '')
learner.load_state(conf, prefix, True, True)
learner.model.eval()
logging.info('learner loaded')
timer.since_last_check('start')
db = Database(dbname)
for ind, res in tqdm.tqdm(enumerate(loader), ):
img = res['img']
imp = res['imgp']
# print(img.shape[0])
sid = res['sid']
tid = res['tid']
# todo basthc size here msut 1
sid = sid[0].item()
tid = tid[0].item()
finish = res['finish']
# extract fea
start = 0
fea_l = []
norm_l = []
with torch.no_grad():
while start < len(img):
img_now = img[start:start + true_batch_size]
# todo multi gpu: fix dataparallel
fea, norm = learner.model(img_now,
return_norm=True,
normalize=False)
start += true_batch_size
fea_l.append(fea.cpu())
norm_l.append(norm.cpu())
fea = torch.cat(fea_l).numpy()
norm = torch.cat(norm_l).numpy()
## save to db
db[f'fea/{sid}/{tid}'] = fea
db[f'norm/{sid}/{tid}'] = norm
db[f'imp/{sid}/{tid}'] = msgpack_dumps(imp)
# msgpack_loads(db[f'imp/{sid}/{tid}'][...].tolist())
# if ind>10:break
db.close()
def init_learner(conf):
learner = face_learner(conf, True)
learner.threshold = THRESHOLD
if conf.device.type == 'cpu':
learner.load_state(conf, 'cpu_final.pth', True, True)
else:
learner.load_state(conf, 'final.pth', True, True)
learner.model.eval()
print('learner loaded')
return learner
def fn_face_verify_module():
mtcnn = MTCNN()
print('mtcnn loaded')
learner = face_learner(conf, True)
learner.threshold = args.threshold
if conf.device.type == 'cpu':
learner.load_state(conf, 'cpu_final.pth', True, True)
else:
learner.load_state(conf, 'final.pth', True, True)
learner.model.eval()
print('learner loaded')
if args.update:
targets, names = prepare_facebank(conf,
learner.model,
mtcnn,
tta=args.tta)
print('facebank updated')
else:
targets, names = load_facebank(conf)
print('facebank loaded')
isSuccess, frame = cap.read()
if isSuccess:
try:
image = Image.fromarray(frame)
bboxes, faces = mtcnn.align_multi(image, conf.face_limit,
conf.min_face_size)
bboxes = bboxes[:, :
-1] # shape:[10,4],only keep 10 highest possibiity faces
bboxes = bboxes.astype(int)
bboxes = bboxes + [-1, -1, 1, 1] # personal choice
results, score = learner.infer(conf, faces, targets, args.tta)
for idx, bbox in enumerate(bboxes):
if args.score:
frame = draw_box_name(
bbox,
names[results[idx] + 1] + '_{:.2f}'.format(score[idx]),
frame)
else:
frame = draw_box_name(bbox, names[results[idx] + 1], frame)
except:
print('detect error')
cv2.imshow('face Capture', frame)
if args.save:
video_writer.write(frame)
def caculate_distance_image(root_folder,parse_text_file):
conf = get_config(False)
mtcnn = MTCNN()
learner = face_learner(conf, True)
learner.threshold = args.threshold
if conf.device.type == 'cpu':
learner.load_state(conf, True, True)
else:
learner.load_state(conf, True, True)
learner.model.eval()
embedding = []
names = []
folder = os.listdir(root_folder)
for file in folder:
try:
img = os.path.join(root_folder,file)
embs, name = extract_single_image(img, conf, learner.model, mtcnn, tta= args.tta)
embedding.append(embs)
names.append(name)
except:
continue
if os.path.exists(parse_text_file):
pass
else:
with open(parse_text_file,"w") as fs:
fs.write('\n CLASS: {} \n'.format(root_folder.split('/')[-1]))
for i in range(1,len(embedding),1):
try:
embs1 = embedding[0]
embs2 = embedding[i]
_, score = distance(embs1, embs2)
line = 'distance image:{} -> image:{} is:{:.4f} \n'.format(names[0],names[i],score[0])
fs.write(line)
except:
continue
for i in range(2,len(embedding)):
try:
_, score = distance(embedding[1],embedding[i])
line = 'distance image:{} -> image:{} is:{:.4f} \n'.format(names[1],names[i],score[0])
fs.write(line)
except:
continue
def __init__(self):
self.stream_url = "https://demo.bahien.com/live/stream/playlist.m3u8"
self.cap = cv2.VideoCapture(self.stream_url)
self.cap.set(cv2.CAP_PROP_POS_MSEC, random.randint(3, 1000))
self.cap.set(cv2.CAP_PROP_BUFFERSIZE, 2)
self.conf = get_config(False)
self.mtcnn = MTCNN()
self.learner = face_learner(self.conf, True)
self.learner.threshold = THRESHOLD
self.learner.load_state(self.conf, 'final.pth', True, True)
self.learner.model.eval()
print('learner loaded')
self.targets, self.names = load_facebank(self.conf)
print('facebank loaded')
self.net, self.ln = yolo_detection.init_net()
print('yolo model loaded')
self.model = Darknet(config_file)
self.model.load_weights(weight_file)
self.model.cuda()
def __init__(self, register=False):
self.conf = get_config()
self.conf.device = torch.device(
"cuda:0" if torch.cuda.is_available() else "cpu")
self.mtcnn = My_Function_lib.MTCNN()
print('mtcnn loaded')
self.learner = face_learner(self.conf, True)
if self.conf.device.type == 'cpu':
self.learner.load_state(self.conf, 'best.pth', True)
else:
self.learner.load_state(self.conf, 'final.pth', True)
self.learner.model.eval()
print('learner loaded')
if register == False:
self.targets, self.names = load_facebank(self.conf)
print('facebank updated')
self.process_this_frame = True
self.Img = None
def register(user_id):
data_path = Path('data')
save_path = data_path / 'facebank' / user_id
fetch_path = data_path / 'dataset' / user_id
images = load_images_from_folder(fetch_path)
print(images)
if not save_path.exists():
save_path.mkdir()
mtcnn = MTCNN()
count = 0
face_id = user_id
count = 0
for img in images:
frame = img
p = Image.fromarray(frame[..., ::-1])
try:
warped_face = np.array(mtcnn.align(p))[..., ::-1]
cv2.imwrite(
"data/facebank/" + str(face_id) + '/' + str(face_id) + '_' +
str(count) + ".jpg", warped_face)
count += 1
#cv2.imwrite(str(save_path/'{}.jpg'.format(str(datetime.now())[:-7].replace(":","-").replace(" ","-"))), warped_face)
except:
result = {
"_result": "Error",
"_message": "Unable to detect the face"
}
if count == len(images):
result = {
"_result": "success",
"_message": "User Registered Successfully"
}
conf = get_config(False)
learner = face_learner(conf, True)
learner.load_state(conf, 'cpu_final.pth', True, True)
learner.model.eval()
#print('learner loaded')
targets, names = prepare_facebank(conf, learner.model, mtcnn, user_id)
#print('facebank updated')
return result
def run(args, conf):
image_root = args.image_root
retina_face_root = args.retinaface
result_root = args.result_root
seq = os.path.basename(image_root)
verbose = args.verbose
if not osp.exists(result_root):
os.makedirs(result_root, exist_ok=True)
learner = face_learner(conf, True)
learner.threshold = args.threshold
learner.load_state(conf, fixed_str="ir_se50.pth", from_save_folder=False, model_only=True)
learner.model.eval()
############################################################
# 1. Construct face bank only using the confident faces
# from openface
############################################################
face_bank, names, emb_counts = construct_face_bank(seq, retina_face_root, image_root, result_root, learner, verbose)
names, emb_counts = np.array(names), np.array(emb_counts)
print("Sequence: {}, found {} unique faces".format(seq, len(emb_counts)))
def caculate_distace_foler(rootdir, parse_text_file):
conf = get_config(False)
mtcnn = MTCNN()
learner = face_learner(conf, True)
learner.threshold = args.threshold
if conf.device.type == 'cpu':
learner.load_state(conf, True, True)
else:
learner.load_state(conf, True, True)
learner.model.eval()
embedding = []
names = []
folder = os.listdir(rootdir)
for path in tqdm.tqdm(folder):
dirs = os.path.join(rootdir,path)
try:
embs, name = extract_folder_image(Path(dirs) ,conf , learner.model, mtcnn, tta= args.tta)
embedding.append(embs)
names.append(name)
except:
continue
# if os.path.exists(parse_text_file):
# pass
# else:
with open(parse_text_file,"w") as fs:
for i in tqdm.tqdm(range(0,len(embedding)-1)):
for j in range(i+1,len(embedding)):
_, score = distance(embedding[i],embedding[j])
print(score)
line = 'distance class:{} -> class:{} is:{:.2f}\n'.format(names[i],names[j],score[0])
fs.write(line)
fs.write(' ===================== \n')
if args.seed is not None:
import numpy as np
import torch
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.cuda.manual_seed(args.seed)
transforms = transforms.get_test_transforms_v2(args.input_size, use_crop=args.use_crop,
center_crop_ratio=args.center_crop_ratio, use_gray=args.use_gray)
dataset = CustomDataset(args.test_img_dir, transforms)
dataloader = torch.utils.data.DataLoader(dataset, batch_size=args.batch_size,
shuffle=False, num_workers=args.num_workers, pin_memory=args.pin_memory)
learner = face_learner(conf, inference=True)
for imgs, labels in self.loader:
imgs = imgs.to(conf.device)
labels = labels.to(conf.device)
self.optimizer.zero_grad()
embeddings = self.model(imgs)
thetas = self.head(embeddings, labels)
device = 'cuda'
total_scores = []
total_indices = []
total_file_names = []
for step, (imgs, file_names) in enumerate(dataloader):
if step > 0 and step % args.log_step_interval == 0:
print(step, len(dataloader))
parser.add_argument("-net", "--net_mode", help="which network, [ir, ir_se, mobilefacenet]",default='ir_se', type=str)
parser.add_argument("-depth", "--net_depth", help="how many layers [50,100,152]", default=50, type=int)
parser.add_argument('-lr','--lr',help='learning rate',default=1e-3, type=float)
parser.add_argument("-b", "--batch_size", help="batch_size", default=96, type=int)
parser.add_argument("-w", "--num_workers", help="workers number", default=3, type=int)
parser.add_argument("-d", "--data_mode", help="use which database, [vgg, ms1m, emore, concat]", default='emore', type=str)
# for continuing training from a trained model
parser.add_argument("-lo", "--load_model",
help="whether model should be loaded before training, enter fixed part of file string",
default=None, type=str)
parser.add_argument("-st", "--step",
help="the starting value for step variable which is used in model name when saving",
default=0, type=int)
args = parser.parse_args()
conf = get_config()
if args.net_mode == 'mobilefacenet':
conf.use_mobilfacenet = True
else:
conf.net_mode = args.net_mode
conf.net_depth = args.net_depth
conf.lr = args.lr
conf.batch_size = args.batch_size
conf.num_workers = args.num_workers
conf.data_mode = args.data_mode
learner = face_learner(conf)
learner.train(conf, args.epochs, fixed_str=args.load_model, step=args.step)
conf.num_sphere = args.num_sphere
if args.pretrain:
conf.pretrain = True
assert len(args.pretrained_model_path)
conf.pretrained_model_path = args.pretrained_model_path
conf.drop_ratio = args.drop_ratio
conf.m = args.m
conf.lr = args.lr
conf.work_path = Path(args.work_path)
conf.model_path = conf.work_path / 'models'
conf.log_path = conf.work_path / 'log'
conf.save_path = conf.work_path / 'save'
if os.path.exists(conf.work_path):
pass
else:
os.mkdir(conf.work_path)
os.mkdir(conf.model_path)
os.mkdir(conf.save_path)
conf.batch_size = args.batch_size
conf.num_workers = args.num_workers
conf.data_mode = args.data_mode
# if conf.data_mode == 'webface':
# conf.epochs = 55
# conf.milestones = [30, 38, 45]
learner = face_learner(conf, embedding_size=512)
learner.train(conf, args.epochs)
default=196,
type=int)
parser.add_argument("-w",
"--num_workers",
help="workers number",
default=8,
type=int)
parser.add_argument(
"-d",
"--data_mode",
help="use which database, [vgg, ms1m, emore, concat, glintasia]",
default='emore',
type=str)
args = parser.parse_args()
conf = get_config()
if args.net_mode == 'seesawFaceNet':
conf.seesawFaceNet = True
else:
conf.net_mode = args.net_mode
conf.net_depth = args.net_depth
conf.lr = args.lr
conf.batch_size = args.batch_size
conf.num_workers = args.num_workers
conf.data_mode = args.data_mode
learner = face_learner(conf, pretrained=True)
learner.train(conf, args.epochs)
default=3,
type=int)
parser.add_argument("-d",
"--data_mode",
help="use which database, [vgg, ms1m, emore, concat]",
default='emore',
type=str)
parser.add_argument("--local_rank", default=0, type=int)
args = parser.parse_args()
print(args.local_rank)
conf = get_config()
torch.distributed.init_process_group(backend='nccl')
torch.cuda.set_device(args.local_rank)
if args.net_mode == 'mobilefacenet':
conf.use_mobilfacenet = True
else:
conf.net_mode = args.net_mode
conf.net_depth = args.net_depth
conf.lr = args.lr
conf.batch_size = args.batch_size
conf.num_workers = args.num_workers
conf.data_mode = args.data_mode
learner = face_learner(conf, args)
t1 = time.time()
learner.train(conf, args.epochs)
t2 = time.time()
print('training time: ', t2 - t1)
def main(args):
start_time = time.time()
merge_datasets = args.merge.split(',')
# load dataset1 class_num & img_size
prop = load_property(merge_datasets[0])
image_size = prop.image_size
print('image_size', image_size)
learner = None
conf = get_config()
# load model
if args.model:
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
print('loading', args.model)
learner = face_learner(conf, train=False)
learner.load_state(conf, '{}.pth'.format(args.model), model_only=True) # r100
learner.model.eval()
else:
print('model is empty')
assert args.t==0.0
rec_list = []
# read rec
for ds in merge_datasets:
path_imgrec = os.path.join(ds, 'train.rec')
path_imgidx = os.path.join(ds, 'train.idx')
imgrec = mx.recordio.MXIndexedRecordIO(path_imgidx, path_imgrec, 'r') # pylint: disable=redefined-variable-type
rec_list.append(imgrec)
id_list_map = {}
all_id_list = []
#
for ds_id in range(len(rec_list)):
id_list = []
imgrec = rec_list[ds_id]
s = imgrec.read_idx(0)
header, _ = mx.recordio.unpack(s)
assert header.flag>0
print('header0 label', header.label)
seq_identity = range(int(header.label[0]), int(header.label[1]))
pp=0
for identity in seq_identity:
pp+=1
if pp%10==0:
print('processing id', pp)
if learner is not None:
embedding = get_embedding(args, imgrec, identity, image_size, learner.model, conf)
else:
embedding = None
#print(embedding.shape)
id_list.append( [ds_id, identity, embedding] )
id_list_map[ds_id] = id_list
if ds_id==0 or learner is None:
all_id_list += id_list
print(ds_id, len(id_list))
else:
X = []
#data_hist = []
for id_item in all_id_list:
X.append(id_item[2])
X = np.array(X)
i = 0
for i in range(len(id_list)):
id_item = id_list[i]
y = id_item[2]
sim = np.dot(X, y.T) #memory error
print(i)
#data_hist.append(sim)
idx = np.where(sim>=args.t)[0]
if len(idx)>0:
continue
all_id_list.append(id_item)
i += 1
#rng_hist = np.arange(-1, 1.05, 0.05)
#data_hist = np.asarray(data_hist)
#res_hist, res_bins, _ = plt.hist(data_hist, rng_hist, rwidth = 0.8)
#print(res_hist)
#plt.show()
if not os.path.exists(args.output):
os.makedirs(args.output)
writer = mx.recordio.MXIndexedRecordIO(os.path.join(args.output, 'train.idx'), os.path.join(args.output, 'train.rec'), 'w')
idx = 1
identities = []
nlabel = -1
for id_item in all_id_list:
if id_item[1]<0:
continue
nlabel+=1
ds_id = id_item[0]
imgrec = rec_list[ds_id]
id = id_item[1]
s = imgrec.read_idx(id)
header, _ = mx.recordio.unpack(s)
a, b = int(header.label[0]), int(header.label[1])
identities.append( (idx, idx+b-a) )
for _idx in range(a,b):
s = imgrec.read_idx(_idx)
_header, _content = mx.recordio.unpack(s)
nheader = mx.recordio.IRHeader(0, nlabel, idx, 0)
s = mx.recordio.pack(nheader, _content)
writer.write_idx(idx, s)
idx+=1
id_idx = idx
for id_label in identities:
_header = mx.recordio.IRHeader(1, id_label, idx, 0)
s = mx.recordio.pack(_header, b'')
writer.write_idx(idx, s)
idx+=1
_header = mx.recordio.IRHeader(1, (id_idx, idx), 0, 0)
s = mx.recordio.pack(_header, b'')
writer.write_idx(0, s)
with open(os.path.join(args.output, 'property'), 'w') as f:
f.write("%d,%d,%d"%(len(identities), image_size[0], image_size[1]))
print('time : {}'.format(time.time()-start_time))
def test(self,conf,img_dir,update=False,view_score=False,view_error=False):
#Load models
mtcnn = MTCNN()
learner = face_learner(conf, True)
if conf.device.type == 'cpu':
learner.load_state(conf,'cpu_final.pth',True,True)
else:
learner.load_state(conf,'final.pth',True,True)
learner.model.eval()
#Load Facebank
if update:
targets, names = prepare_facebank(conf, learner.model, mtcnn, False)
print('facebank updated')
else:
targets, names = load_facebank(conf)
print('facebank loaded')
#Load Image list
img_list = glob(img_dir + '**/*.jpg')
acc = 0
detect_err=0
fails = []
print(f"{'Found':^15}{'Name':^20}{'Result':^15}{'Score':^15}")
pbar = enumerate(img_list)
pbar = tqdm(pbar, total = len(img_list))
for i, x in pbar:
preds = []
label = str(os.path.dirname(x))
label = os.path.basename(label)
image = Image.open(x)
frame = cv2.imread(x,cv2.IMREAD_COLOR)
frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
try:
bboxes, faces = mtcnn.align_multi(image, conf.face_limit, conf.min_face_size)
bboxes = bboxes[:,:-1] #shape:[10,4],only keep 10 highest possibiity faces
bboxes = bboxes.astype(int)
bboxes = bboxes + [-1,-1,1,1] # personal choice
results, score = learner.infer(conf, faces, targets, False)
for idx,bbox in enumerate(bboxes):
print(f'{Label}: {score[idx]}')
if view_score:
frame = draw_box_name(bbox, names[results[idx] + 1] + '_{:.2f}'.format(score[idx]), frame)
else:
frame = draw_box_name(bbox, names[results[idx] + 1], frame)
preds.append(names[results[idx]+1])
if label in preds:
acc += 1
else:
fails.append([label,preds])
# Image.fromarray(frame,'RGB').show()
except Exception as ex:
fails.append([label,ex])
detect_err += 1
f = len(bboxes)
tf = str(True if label in preds else False)
t = f'{f:^15}{label:^20}{tf:^15}{acc/(i+1):^15.4}'
pbar.set_description(t)
if detect_err>0:
print(f'Detect Error: {detect_err}')
if view_error:
pp(fails)
else:
print(f'If you want to see details, make veiw_error True.')
print(f'Accuracy: {acc/len(img_list)}')
def predict():
#img=img_path#注意修改
parser = argparse.ArgumentParser(description='for face verification')
parser.add_argument("-s",
"--save",
help="whether save",
action="store_true")
parser.add_argument('-th',
'--threshold',
help='threshold to decide identical faces',
default=1.54,
type=float)
parser.add_argument("-u",
"--update",
default=True,
help="whether perform update the facebank",
action="store_true")
parser.add_argument("-tta",
"--tta",
default=False,
help="whether testCode time augmentation",
action="store_true")
parser.add_argument("-c",
"--score",
default=True,
help="whether show the confidence score",
action="store_true")
parser.add_argument('--img_path',
'-p',
default='1.jpg',
type=str,
help='input the name of the recording person')
args = parser.parse_args()
mtcnn = MTCNN()
conf = get_config(False)
learner = face_learner(conf, True)
learner.threshold = args.threshold
image = Image.open(
'D:/code/Python/InsightFace-刘悦/InsightFace_Pytorch-master/1.jpg')
if conf.device.type == 'cpu':
learner.load_state(conf, 'ir_se50.pth', True, True)
else:
learner.load_state(conf, 'ir_se50.pth', True, True)
learner.model.eval()
print('learner loaded')
if args.update:
targets, names = prepare_facebank(conf,
learner.model,
mtcnn,
tta=args.tta)
print('facebank updated')
else:
targets, names = load_facebank(conf)
print('facebank loaded')
try:
# image = Image.fromarray(img)
# 利用mtcnn网络，对齐
bboxes, faces = mtcnn.align_multi(image, conf.face_limit,
conf.min_face_size)
bboxes = bboxes[:, :
-1] # shape:[10,4],only keep 10 highest possibiity faces
bboxes = bboxes.astype(int)
bboxes = bboxes + [-1, -1, 1, 1] # personal choice
results, score = learner.infer(conf, faces, targets, args.tta)
print(results, score)
for idx, bbox in enumerate(bboxes):
print(names[results[idx] + 1])
res = "name:" + names[results[idx] + 1]
except:
print('detect')
res = "unknow"
return res
if args.benchmark:
import torch.backends.cudnn as cudnn
cudnn.benchmark = True
cudnn.deterministic = True
train_transform_func = getattr(transforms, args.train_transform_func_name)
train_transforms = train_transform_func(
args.input_size,
args.use_random_crop,
args.use_gray,
only_use_pixel_transform=args.only_use_pixel_transform,
use_flip=args.use_flip,
use_blur=args.use_blur,
no_transforms=args.no_transforms,
use_center_crop=args.use_center_crop,
center_crop_ratio=args.center_crop_ratio)
val_transform_func = getattr(transforms, args.val_transform_func_name)
val_transforms = val_transform_func(
input_size=args.input_size,
use_gray=args.use_gray,
use_center_crop=args.use_center_crop,
center_crop_ratio=args.center_crop_ratio)
learner = face_learner(conf,
train_transforms=train_transforms,
val_transforms=val_transforms)
learner.train(conf, args.epochs)
def train(self,conf):
print(f'Train Start. Train dataset is {conf.data_mode}')
learner = face_learner(conf,False)
learner.train(conf, conf.epoch)
conf.ipabn = False
conf.cvt_ipabn = False
conf.fill_cache = False
# conf.net_depth = 152
# conf.net_mode = 'mobilefacenet'
conf.use_chkpnt = False
from Learner import FaceInfer, face_learner
# learner = FaceInfer(conf, gpuid=range(conf.num_devs))
# learner.load_state(
# resume_path=f'work_space/{args.modelp}/models/',
# latest=False,
# )
# learner.model.eval()
learner = face_learner()
learner.load_state(
resume_path=f'work_space/{args.modelp}/models/',
latest=False,
load_optimizer=False,
load_imp=False,
load_head=False,
)
learner.model.eval()
from pathlib import Path
res = {}
for ds in [
'agedb_30',
'lfw',
def face_detection(video_path, result_csv_path):
conf = get_config(False)
mtcnn = MTCNN(select_largest=False, keep_all=True)
print('mtcnn loaded')
learner = face_learner(conf, True)
learner.threshold = 0.5
if conf.device.type == 'cpu':
learner.load_state(conf, 'cpu_final.pth', True, True)
else:
learner.load_state(conf, 'final.pth', True, True)
learner.model.eval()
print('learner loaded')
targets, names = prepare_facebank(conf, learner.model, mtcnn, tta = True)
print('facebank updated')
cap = cv2.VideoCapture(video_path)
cap.set(cv2.CAP_PROP_POS_MSEC, 0)
fps = cap.get(cv2.CAP_PROP_FPS)
#video writer
#video_writer = cv2.VideoWriter(str(conf.facebank_path/'{}.avi'.format(args.save_name)),
# cv2.VideoWriter_fourcc(*'XVID'), int(fps), (int(cap.get(3)),int(cap.get(4))))
#video_writer = cv2.VideoWriter(str(conf.facebank_path/'{}.avi'.format(args.save_name)),
# cv2.CAP_OPENCV_MJPEG, int(fps), (int(cap.get(3)),int(cap.get(4))))
#### csv
df = pd.DataFrame(columns=['frame_number', 'ID', 'LT', 'RB', 'score'])
framecounter = 0
while cap.isOpened():
isSuccess,frame = cap.read()
if isSuccess:
# image = Image.fromarray(frame[...,::-1]) #bgr to rgb
image = Image.fromarray(frame)
try:
bboxes, faces = mtcnn.align_multi(image, conf.face_limit, 16)
# mtcnn 에서 검출된 얼굴 BB 와 5point landmark
except:
bboxes = []
faces = []
if len(bboxes) == 0:
print('no face')
# continue
else:
bboxes = bboxes[:,:-1] #shape:[10,4],only keep 10 highest possibiity faces
bboxes = bboxes.astype(int)
bboxes = bboxes + [-1,-1,1,1] # personal choice
# 사람 identification
results, score = learner.infer(conf, faces, targets, True)
for idx,bbox in enumerate(bboxes):
#frame = draw_box_name(bbox, names[results[idx] + 1] + '_{:.2f}'.format(score[idx]), frame)
df = df.append({'frame_number':framecounter, 'ID':names[results[idx] + 1], 'LT':(bbox[0],bbox[1]), 'RB':(bbox[2],bbox[3]), 'score':'{:.2f}'.format(score[idx])}, ignore_index=True)
#video_writer.write(frame)
#print('{0:.2f}' .format(framecounter/duration*100))
framecounter +=1
else:
break
cap.release()
#video_writer.release()
df.to_csv(result_csv_path, index=False)
import pims, cvbase as cvb
from lz import *
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='for face verification')
parser.add_argument("-s", "--save", help="whether save", action="store_true", default=True)
parser.add_argument('-th', '--threshold', help='threshold to decide identical faces', default=1.54, type=float)
parser.add_argument("-u", "--update", help="whether perform update the facebank", action="store_true",
default=False)
parser.add_argument("-tta", "--tta", help="whether test time augmentation", action="store_true", default=True)
parser.add_argument("-c", "--score", help="whether show the confidence score", action="store_true", default=True)
args = parser.parse_args()
conf = get_config(False)
learner = face_learner(conf, True)
learner.threshold = args.threshold
# conf.work_path = Path('work_space/')
if conf.device.type == 'cpu':
learner.load_state(conf, 'cpu_final.pth', True, True)
else:
learner.load_state(conf, 'final.pth', True, True)
learner.model.eval()
print('learner loaded')
def extract_fea_from_img(img):
img = img.copy()[..., ::-1].reshape(112, 112, 3)
img = Image.fromarray(img)
mirror = transforms.functional.hflip(img)
with torch.no_grad():
def main1():
parser = argparse.ArgumentParser(description='for face verification')
parser.add_argument("-e",
"--epochs",
help="training epochs",
default=8,
type=int)
parser.add_argument("-net",
"--net_mode",
help="which network, [ir, ir_se, mobilefacenet]",
default='ir_se',
type=str)
parser.add_argument("-depth",
"--net_depth",
help="how many layers [50,100,152]",
default=50,
type=int)
parser.add_argument('-lr',
'--lr',
help='learning rate',
default=1e-3,
type=float)
parser.add_argument("-b",
"--batch_size",
help="batch_size",
default=96,
type=int)
parser.add_argument("-w",
"--num_workers",
help="workers number",
default=3,
type=int)
parser.add_argument("-d",
"--data_mode",
help="use which database, [vgg, ms1m, emore, concat]",
default='emore',
type=str)
parser.set_defaults(
epochs=8,
net='ir_se',
net_depth='50',
lr=1e-3,
batch_size=96,
num_workers=3,
data_mode="ms1m",
)
args = parser.parse_args()
conf = get_config(training=True)
if args.net_mode == 'mobilefacenet':
conf.use_mobilfacenet = True
else:
conf.net_mode = args.net_mode
conf.net_depth = args.net_depth
conf.lr = args.lr
conf.batch_size = args.batch_size
conf.num_workers = args.num_workers
conf.data_mode = args.data_mode
learner = face_learner(conf, inference=False, need_loader=False)
# print(learner.find_lr(conf, ))
# learner.train(conf, args.epochs)
for i in range(1):
for imgs, labels in learner.loader:
imgs = imgs.cuda()
labels = labels.cuda()
print('ok', imgs.shape, labels.shape)
embeddings = learner.model(imgs)
thetas = learner.head(embeddings, labels)
loss = conf.ce_loss(thetas, labels)
import torch.autograd
# fgg g
grad = torch.autograd.grad(loss,
embeddings,
retain_graph=False,
create_graph=False,
only_inputs=True)[0].detach()
embeddings_adv = embeddings + 0.01 * grad
thetas_adv = learner.head(embeddings_adv, labels)
loss_adv = conf.ce_loss(thetas_adv, labels)
loss_adv.backward()
def main():
args = parser.parse_args()
print('args.data_url', args.data_url)
if conf.cloud:
mox.file.copy_parallel(args.data_url, '/cache/face_train/')
args.data_url = '/cache/face_train/'
conf.use_data_folder = args.data_url
if args.work_path:
conf.work_path = Path(args.work_path)
conf.model_path = conf.work_path / 'models'
conf.log_path = conf.work_path / 'log'
conf.save_path = conf.work_path / 'save'
else:
args.work_path = conf.work_path
conf.update(args.__dict__)
if conf.local_rank is not None:
torch.cuda.set_device(conf.local_rank)
torch.distributed.init_process_group(backend='nccl',
init_method="env://")
if torch.distributed.get_rank() != 0:
set_stream_logger(logging.WARNING)
# if osp.exists(conf.save_path):
# logging.info('ok')
# exit(1)
# simplify_conf(conf)
# exit(0)
from Learner import face_learner
# decs = msgpack_load('decs.pk')
# conf.decs = decs
learner = face_learner(conf, )
# fstrs = learner.list_fixed_strs('work_space/sglpth.casia/models')
# stps = learner.list_steps('work_space/sglpth.casia/models')
# fstr = fstrs[np.argmax(stps)]
# stt_dct = torch.load('work_space/sglpth.casia/models/model_' + fstr)
# learner.model.module.load_state_dict_sglpth(stt_dct)
# print(fstrs, stps, fstr, )
if conf.get('load_from'):
# p= 'r100.128.retina.clean.arc',
# 'hrnet.retina.arc.3',
# 'mbv3.retina.arc',
# 'mbfc.lrg.retina.arc.s48',
# 'effnet.casia.arc',
# 'mbfc.retina.cl.distill.cont2',
# 'mbfc2',
# 'r18.l2sft',
# 'r18.adamrg',
# 'mbfc.se.elu.ms1m.radam.1',
# 'mbfc.se.elu.specnrm.allbutdw.ms1m.adam.1',
# 'mbfc.se.prelu.specnrm.ms1m.cesigsft.1',
# 'irse.elu.ms1m',
# 'irse.elu.casia.arc.2048',
p = Path(conf.load_from)
print(
'try to load from ',
p,
)
learner.load_state(
resume_path=p,
load_optimizer=False,
load_head=conf.head_load, # todo note!
load_imp=False,
latest=True,
strict=False,
)
# simplify_conf(conf)
learner.cloud_sync_log()
# res = learner.validate_ori(conf, valds_names=('cfp_fp', ))
# exit(0)
# learner.calc_img_feas(out='work_space/mbfc.crash.h5')
# log_lrs, losses = learner.find_lr(
# num=999,
# bloding_scale=1000)
# losses[np.isnan(losses)] = 999
# best_lr = 10 ** (log_lrs[np.argmin(losses)])
# print('best lr is ', best_lr)
# conf.lr = best_lr
# exit(0)
# learner.init_lr()
# conf.tri_wei = 0
# log_conf(conf)
# learner.train(conf, 1, name='xent')
learner.init_lr()
simplify_conf(conf)
if conf.head_init:
learner.head_initialize()
if conf.warmup:
learner.warmup(conf, conf.warmup)
learner.train_simple(conf, conf.epochs)
# learner.train_dist(conf, conf.epochs)
if conf.net_mode == 'sglpth':
decs = learner.model.module.get_decisions()
msgpack_dump(decs, 'decs.pk')
# learner.train_cotching(conf, conf.epochs)
# learner.train_cotching_accbs(conf, conf.epochs)
# learner.train_ghm(conf, conf.epochs)
# learner.train_with_wei(conf, conf.epochs)
# learner.train_use_test(conf, conf.epochs)
# res = learner.validate_ori(conf, )
if not conf.cloud:
from tools.test_ijbc3 import test_ijbc3
res = test_ijbc3(conf, learner)
tpr6, tpr4, tpr3 = res[0][1], res[1][1], res[2][1]
learner.writer.add_scalar('ijbb/6', tpr6, learner.step)
learner.writer.add_scalar('ijbb/4', tpr4, learner.step)
learner.writer.add_scalar('ijbb/3', tpr3, learner.step)
learner.writer.close()
if conf.never_stop:
img = torch.randn((conf.batch_size // 2, 3, conf.input_size,
conf.input_size)).cuda()
learner.model.eval()
logging.info('never stop')
while True:
_ = learner.model(img)
def update_facebank(self,conf):
mtcnn = MTCNN()
learner = face_learner(conf,True)
learner.load_state(conf,'final.pth',True,True)
learner.model.eval()
_, _ = prepare_facebank(conf,learner.model,mtcnn,False)