def __init__(self,db_file,id_dir,save_dir,label_file=None):
if faceconfig.torch_:
model_path = "../models/torch_models/resnet50m-xent-face1.pth.tar"
FaceModel = Deep_Face(model_path,256,128)
elif faceconfig.mx_:
model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-r100-ii/model"
#model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-r50-am-lfw/model"
#model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-r34-amf/model"
#model_path = "../models/mx_models/v1_bn/model" #9
#model_path = "../models/mx_models/model-100/model-v3/modelresave"
epoch_num = 0 #9
img_size = [112,112]
FaceModel = mx_Face(model_path,epoch_num,img_size)
elif faceconfig.caffe_use:
face_p = "../models/sphere/sph_2.prototxt"
face_m = "../models/sphere/sph20_ms_4v5.caffemodel"
#face_p2 = "../models/mx_models/mobile_model/face-mobile.prototxt"
#face_m2 = "../models/mx_models/mobile_model/face-mobile.caffemodel"
if config.feature_1024:
out_layer = 'fc5_n'
elif config.insight:
out_layer ='fc1'
else:
out_layer ='fc5'
FaceModel = FaceReg(face_p,face_m,out_layer)
self.FaceModel = FaceModel
self.label_file = label_file
self.id_dir = id_dir
self.DB_FT = Annoy_DB(faceconfig.feature_lenth)
self.db_file = db_file
self.save_dir = save_dir
self.mkdb()
self.db_cnt_dict = dict()
self.dist_path = None
self.img_name = None
self.img_cnt = 0
def Db_test(label_file, **kwargs):
#db_file,data_file,base_dir,id_dir,dis_dir,base_id
db_file = kwargs.get('db_file', None)
data_file = kwargs.get('data_file', None)
base_dir = kwargs.get('base_dir', None)
id_dir = kwargs.get('id_dir', None)
dis_dir = kwargs.get('saved_dir', None)
base_id = kwargs.get('base_id', None)
failed_dir = kwargs.get('failed_dir', None)
db_in = open(db_file, 'r')
data_in = open(data_file, 'r')
db_lines = db_in.readlines()
data_lines = data_in.readlines()
dis_record = open("./output/distance_top1.txt", 'w')
dis_top2 = open("./output/distance_top2.txt", 'w')
if config.caffe_use:
face_p1 = "../models/sphere/sph_2.prototxt"
face_m1 = "../models/sphere/sph20_ms_4v5.caffemodel"
face_p2 = "../models/mx_models/mobile_model/face-mobile.prototxt"
face_m2 = "../models/mx_models/mobile_model/face-mobile.caffemodel"
if config.feature_1024:
out_layer = 'fc5_n'
elif config.insight:
out_layer = 'fc1'
else:
out_layer = 'fc5'
FaceModel_1 = FaceReg(face_p1, face_m1, 'fc5')
FaceModel_2 = FaceReg(face_p2, face_m2, out_layer)
elif config.mx_:
#model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-r50-am-lfw/model"
#model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-r34-amf/model"
#model_path = "../models/mx_models/v1_bn/model" #9
#model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-r100-ii/model"
#model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-y1-test2/model"
#model_path = "../models/mx_models/mobile_model/model"
model_path1 = "../models/mx_models/model-100/model-org/modelresave"
#model_path = "../models/mx_models/model_prison/model"
#model_path = "../models/mx_models/model-r50/model"
#model_path = "../models/mx_models/model-100/model-v2/model"
model_path2 = "../models/mx_models/model-100/model-v3/modelresave"
epoch1_num = 0 #9#2
epoch2_num = 0
img_size = [112, 112]
FaceModel_1 = mx_Face(model_path1, epoch1_num, img_size)
FaceModel_2 = mx_Face(model_path2, epoch2_num, img_size)
else:
print("please select a frame: caffe mxnet tensorflow")
#mkdir save image dir
Face1_features = ExtractFeatures(FaceModel_1)
Face2_features = ExtractFeatures(FaceModel_2)
make_dirs(dis_dir)
make_dirs(failed_dir)
parm = args()
DB_FT = Annoy_DB(config.feature_lenth)
tpr = 0
fpr = 0
idx_ = 0
db_names = []
save_reg_dirs = []
db_cnt_dict = dict()
db_paths_dict = dict()
#db_features = []
for item_, line_one in enumerate(db_lines):
line_one = line_one.strip()
base_feat1 = Face1_features.extract_f(line_one, id_dir)
base_feat2 = Face2_features.extract_f(line_one, id_dir)
if base_feat1 is None or base_feat2 is None:
print("feature is None")
continue
base_feat = featuresProcess([base_feat1, base_feat2])
one_name = line_one[:-4]
db_names.append(one_name)
face_db_dir = os.path.join(dis_dir, one_name)
save_reg_dirs.append(face_db_dir)
DB_FT.add_data(item_, base_feat)
db_path = os.path.join(id_dir, line_one)
db_paths_dict[one_name] = db_path
#db_features.append(base_feat)
#db_features = np.asarray(db_features)
print("begin do build db")
#print(db_features.shape)
#DB_FT = Face_DB(db_features)
DB_FT.build_db()
print("db over")
label_dict = get_label(label_file)
dest_id_dict = dict()
org_id_dict = dict()
org_wrong_dict = dict()
cnt_org = 0
cnt_dest = 0
test_label_dict = dict()
reg_frame_dict = dict()
for querry_line in data_lines:
querry_line = querry_line.strip()
if config.use_framenum:
frame_num = get_frame_num(querry_line)
if config.reg_fromlabel:
que_spl = querry_line.split("/")
if config.dir_label:
key_que = que_spl[0]
else:
key_que = que_spl[0][2:]
test_label_dict[key_que] = 1
real_label = label_dict.setdefault(key_que, 300) + base_id
idx_ += 1
sys.stdout.write('\r>> deal with %d/%d' % (idx_, len(data_lines)))
sys.stdout.flush()
if config.face_detect:
querry_feat1 = Face1_features.extract_f3(querry_line, base_dir)
querry_feat2 = Face2_features.extract_f3(querry_line, base_dir)
else:
querry_feat1 = Face1_features.extract_f2(querry_line, base_dir)
querry_feat2 = Face2_features.extract_f2(querry_line, base_dir)
if querry_feat1 is None or querry_feat2 is None:
print("feature is None")
continue
else:
querry_feat = featuresProcess([querry_feat1, querry_feat2])
idx_list, distance_list = DB_FT.findNeatest(querry_feat, 2)
#print("distance, idx ",distance,idx)
idx = idx_list[0]
img_name = db_names[idx]
img_dir = save_reg_dirs[idx]
pred_label = label_dict.setdefault(img_name, 300) + base_id
img_cnt = db_cnt_dict.setdefault(img_name, 0)
org_path = os.path.join(base_dir, querry_line)
Threshold_Value = config.confidence
reg_condition = 1 if (distance_list[0] <= config.top1_distance and
(distance_list[1] -
distance_list[0]) >= Threshold_Value) else 0
if config.use_framenum:
if reg_condition:
reg_frame_cnt = reg_frame_dict.setdefault(
img_name, frame_num)
frame_interval = np.abs(frame_num - reg_frame_cnt)
save_condition = 1 if frame_interval < config.frame_interval else 0
else:
save_condition = 0
else:
save_condition = reg_condition
#dist_path = os.path.join(img_path,img_name+"_"+str(img_cnt)+".jpg")
#if config.reg_fromlabel:
# dist_path = os.path.join(img_dir,que_spl[1])
#else:
# dist_path = os.path.join(img_dir,querry_line)
#print(dist_path)
print("distance ", distance_list)
dis_record.write("%.3f\n" % (distance_list[0]))
dis_top2.write("%.3f\n" % (distance_list[1] - distance_list[0]))
#if distance_list[0] <= config.top1_distance and (distance_list[1] - distance_list[0]) >= Threshold_Value:
if save_condition:
print("distance: ", distance_list[0])
make_dirs(img_dir)
if config.reg_fromlabel:
dist_path = os.path.join(img_dir, que_spl[1])
else:
dist_path = os.path.join(img_dir, querry_line)
shutil.copyfile(org_path, dist_path)
db_cnt_dict[img_name] = img_cnt + 1
if db_cnt_dict[img_name] == 1 and config.save_idimg:
org_id_path = db_paths_dict[img_name]
dist_id_path = os.path.join(img_dir, img_name + ".jpg")
shutil.copyfile(org_id_path, dist_id_path)
if config.reg_fromlabel:
print("real and pred ", real_label, pred_label,
querry_line)
if int(pred_label) == int(real_label):
tpr += 1
print("*************")
else:
fpr += 1
if config.save_failedimg:
failed_img_dir = os.path.join(failed_dir, key_que)
make_dirs(failed_img_dir)
failed_img_path = os.path.join(
failed_img_dir, que_spl[1])
shutil.copyfile(org_path, failed_img_path)
cnt_org = org_id_dict.setdefault(key_que, 0)
org_id_dict[key_que] = cnt_org + 1
cnt_dest = dest_id_dict.setdefault(img_name, 0)
dest_id_dict[img_name] = cnt_dest + 1
ori_dest_id = org_wrong_dict.setdefault(key_que, [])
if img_name in ori_dest_id:
pass
else:
org_wrong_dict[key_que].append(img_name)
else:
print("pred label: ", pred_label, querry_line)
tpr += 1
print("*************")
elif config.save_failedimg and not config.reg_fromlabel:
failed_img_dir = os.path.join(failed_dir, img_name)
make_dirs(failed_img_dir)
failed_img_path = os.path.join(failed_img_dir, querry_line)
shutil.copyfile(org_path, failed_img_path)
right_id = 0
not_reg_id = []
for key_name in db_cnt_dict.keys():
if db_cnt_dict[key_name] != 0:
right_id += 1
elif config.reg_fromlabel:
if key_name in test_label_dict.keys():
not_reg_id.append(key_name)
else:
not_reg_id.append(key_name)
print("not reg keys: ", not_reg_id)
print("right id ", right_id)
if config.reg_fromlabel:
org_id = 0
#ori_name = []
for key_name in org_id_dict.keys():
if org_id_dict[key_name] != 0:
org_id += 1
#ori_name.append[key_name]
dest_id = 0
#dest_name = []
for key_name in dest_id_dict.keys():
if dest_id_dict[key_name] != 0:
dest_id += 1
#dest_name.append[key_name]
f_result = open("./output/result_record.txt", 'w')
for key_name in org_wrong_dict.keys():
f_result.write("{} : {}".format(key_name,
org_wrong_dict[key_name]))
f_result.write("\n")
#print("values: ",db_cnt_dict.values())
f_result.close()
print("TPR and FPR is ", tpr, fpr)
print("wrong orignal: ", org_id_dict.values())
print("who will be wrong: ", dest_id_dict.values())
print("the org and dest: ", org_id, dest_id)
else:
print("Reg img num: ", tpr)
db_in.close()
data_in.close()
dis_record.close()
dis_top2.close()
def compare_img(path1, path2, model_path):
parm = args()
img1, img_1 = read_img(path1)
img2, img_2 = read_img(path2)
#model_path = "../models/tf-models/InsightFace_iter-3140"
if config.mx_:
#model_path = "../models/mx_models/v1_bn/model"
#model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-r100-ii/model"
#model_path = "/home/lxy/Develop/Center_Loss/arcface/insightface/models/model-y1-test2/model"
#model_path = "../models/mx_models/mobile_model/model"
model_path1 = "../models/mx_models/model-100/model-v3/modelresave"
model_path2 = "../models/mx_models/model-100/model-v2/model"
epoch1_num = 0 #9#2
epoch2_num = 5
img_size = [112, 112]
FaceModel_1 = mx_Face(model_path1, epoch1_num, img_size)
FaceModel_2 = mx_Face(model_path2, epoch2_num, img_size)
elif config.caffe_use:
face_p1 = "../models/sphere/sph_2.prototxt"
face_m1 = "../models/sphere/sph20_ms_4v5.caffemodel"
face_p2 = "../models/mx_models/mobile_model/face-mobile.prototxt"
face_m2 = "../models/mx_models/mobile_model/face-mobile.caffemodel"
if config.feature_1024:
out_layer = 'fc5_n'
elif config.insight:
out_layer = 'fc1'
else:
out_layer = 'fc5'
FaceModel_1 = FaceReg(face_p1, face_m1, 'fc5')
FaceModel_2 = FaceReg(face_p2, face_m2, out_layer)
elif config.tf_model:
img_size = [112, 112]
FaceModel = TF_Face_Reg(model_path, img_size, parm.gpu)
else:
print("please select a model frame")
def norm_feat(feat):
f_mod = np.sqrt(np.sum(np.power(feat, 2)))
return feat / f_mod
def std_norm(feat):
mean1 = np.mean(feat)
[
lenth,
] = np.shape(feat)
f_center1 = feat - mean1
std1 = np.sum(np.power(f_center1, 2))
std1 = np.sqrt(std1 / lenth)
norm1 = f_center1 / std1
return norm1
if img1 is not None and img2 is not None:
#save_mat(img1,"img1.txt")
if config.caffe_use:
feature_1 = FaceModel.extractfeature(img1)
feat1 = np.array(feature_1)
feature_2 = FaceModel.extractfeature(img2)
feat2 = np.array(feature_2)
elif config.mx_:
feature_1 = FaceModel.extractfeature(img_1)
feat1 = np.array(feature_1)
feature_2 = FaceModel.extractfeature(img_2)
feat2 = np.array(feature_2)
#feat1 = np.array(feature_1)
#np.savetxt("feat1.txt",feat1)
if config.debug:
print("first feat1 ", feat1[:5])
print("first feat2 ", feat2[:5])
#save_mat(img2,"img2.txt")
#feature_2 = FaceModel.extractfeature(img2)
#feat2 = np.array(feature_2)
#np.savetxt("feat2.txt",feat2)
if config.norm:
feat2 = norm_feat(feat2)
feat1 = norm_feat(feat1)
#feat2 = std_norm(feat2)
#feat1 = std_norm(feat1)
dis_s = FaceModel.calculateL2(feat1, feat2)
dis2 = FaceModel.calculateL2(feat1, feat2, 'cosine')
dis3 = FaceModel.calculateL2(feat1, feat2, 'correlation')
if config.debug:
print("feat1 ", feat1[:5])
print("feat2 ", feat2[:5])
print("distance: eucli, cos, cor %.3f,%.3f,%.3f" % (dis_s, dis2, dis3))
def extract_f(model_, img):
#img_path = os.path.join(base_dir,path_)
#img = cv2.imread(img_path)
if img is None:
feat = None
print("read img faild, " + img_path)
else:
if config.face_enhance:
img = Img_enhance(img)
img = cv2.cvtColor(img, cv2.COLOR_BGR2RGB)
if config.torch_:
img = img * 1.0
img = img / 255
elif config.caffe_use:
img = (img - 127.5) * 0.0078125
feat = model_.extractfeature(img)
if config.norm:
#feat = norm_feat(feat)
pass
return feat
feature_1 = extract_f(FaceModel, img_1)
feat1 = np.array(feature_1)
feature_2 = extract_f(FaceModel, img_2)
feat2 = np.array(feature_2)
if config.debug:
print("second feat1 ", feat1[:5])
print("second feat2 ", feat2[:5])
distance_ecud = L2_distance(feat1, feat2, config.feature_lenth)
distance_ecud2 = L2_distance_(feat1, feat2, config.feature_lenth)
print("distance: eucli1,ecu2, %.3f %.3f" %
(distance_ecud, distance_ecud2))
cv2.imshow("img1", img_1)
cv2.imshow("img2", img_2)
cmd = cv2.waitKey(0) & 0xFF
if cmd == 27 or cmd == ord('q'):
cv2.destroyAllWindows()