Training mode

Two training modes are included currently, i.e., conventional training and semi-siamese training. Edit the configuration of each training mode by the following steps, and then you can train a face recognition model by the certain mode.

1. Training data

We use MS-Celeb-1M-v1c for conventional training. To perform open-set evaluation, we try our best to remove the identities which may overlap between this dataset and all of the test sets, resulting in a training set which includes 72,778 identities and about 3.28M images. The final identity list can be found in MS-Celeb-1M-v1c-r_id_list.txt. The format of training list should be the same as MS-Celeb-1M-v1c-r_train_list.txt. The shallow training set MS-Celeb-1M-v1c-Shallow is formed by randomly selecting two images of an identity in MS-Celeb-1M-v1c, and the selected image list can be downloaded in MS-Celeb-1M-v1c-r-shallow_train_list.txt. The training set for masked face recognition(MS-Celeb-1M-v1c-Mask) includes the original face images of each identity in MS-Celeb1M-v1c, as well as the corresponding masked face image by FMA-3D.

2. Train a face recognition model

Step1: Prepare the training data

Align the face images to 112*112 according to face_align.py.

Step2: Configure the backbone

Edit the configuration in backbone_conf.yaml. Detailed description about the configuration can be found in backbone_def.py.

Step3: Configure the head

Edit the configuration in head_conf.yaml. Detailed description about the configuration can be found in head_def.py.

Step4: Configure the training setting for each mode.

Edit the configuration in train.sh. Detailed description about the configuration can be found in train.py.

Step5: Start training

sh train.sh

Other Tips

• In order to train a model using only the upper half of face (model2 in 3.4), you need to set the last parameter of 'ImageDataset' to True and modify the 'out_h' of the backbone to 4.
• In order to train the masked face recognition model (model3 in 3.4), you just need to change the training set to MS-Celeb-1M-v1c-Mask, which includes 72,778 identities and about 3.28*2M images.

3. Trained models and logs

The models and training logs mentioned in our technical report are listed as follows. You can click the link to download them. For Megaface, we report the accuracy of the last checkpoint, and for other benchmarks, we report the accuracy of the best checkpoint.

3.1 Experiments of SOTA backbones

Backbone	LFW	CPLFW	CALFW	AgeDb	MegaFace	Params	Macs	Models&Logs
MobileFaceNet	99.57	83.33	93.82	95.97	90.39	1.19M	227.57M	Google,Baidu:bmpn
Resnet50-ir	99.78	88.20	95.47	97.77	96.67	43.57M	6.31G	Google,Baidu:8ecq
Resnet152-irse	99.85	89.72	95.56	98.13	97.48	71.14M	12.33G	Google,Baidu:2d0c
HRNet	99.80	88.89	95.48	97.82	97.32	70.63M	4.35G	Google,Baidu:t9eo
EfficientNet-B0	99.55	84.72	94.37	96.63	91.38	33.44M	77.83M	Google,Baidu:sgja
TF-NAS-A	99.75	85.90	94.87	97.23	94.42	39.59M	534.41M	Google,Baidu:kq2v
LightCNN-29	99.57	82.60	93.87	95.78	89.32	11.60M	2.84G	Google,Baidu:kq2v
GhostNet	99.65	83.52	93.93	95.70	89.42	26.76M	194.49M	Google,Baidu:6dg1
Attention-56	99.88	89.18	95.65	98.12	97.75	98.96M	6.34G	Google,Baidu:f93u
Attention-92(MX)	99.82	90.33	95.88	98.08	98.09	134.56M	10.62G	Google,Baidu:3ura
ResNeSt50	99.80	89.98	95.55	97.98	97.08	76.79M	5.55G	Google,Baidu:3ura
ReXNet_1.0	99.65	84.68	94.58	96.70	93.17	15.20M	429.64M	Google,Baidu:3ura
RepVGG_A0	99.77	85.43	94.88	96.97	94.40	39.94M	1.55G	Google,Baidu:gdsf
RepVGG_B0	99.72	86.77	95.17	97.57	95.75	46.65M	3.44G	Google,Baidu:ip68
RepVGG_B1	99.82	87.55	95.50	97.78	96.74	106.75M	13.21G	Google,Baidu:b60b
Swin-T	99.87	88.57	95.56	97.90	97.83	46.74M	4.37G	Google,Baidu:17ww
Swin-S	99.85	90.03	95.92	98.05	98.17	68.01M	8.53G	Google,Baidu:hhre

• MegaFace means MegaFace rank1 accuracy.
• Params and Macs are computed by THOP.
• MX means mixed precision training by apex.

3.2 Experiments of SOTA heads

Supervisory Head	LFW	CPLFW	CALFW	AgeDb	MegaFace_rank1	Models&Logs
AM-Softmax	99.58	83.63	93.93	95.85	88.92	Google,Baidu:pe3n
AdaM-Softmax	99.58	83.85	93.50	96.02	89.40	Google,Baidu:rcrk
AdaCos	99.65	83.27	92.63	95.38	82.95	Google,Baidu:3sef
ArcFace	99.57	83.68	93.98	96.23	88.39	Google,Baidu:aujd
MV-Softmax	99.57	83.33	93.82	95.97	90.39	Google,Baidu:fcpd
CurricularFace	99.60	83.03	93.75	95.82	87.27	Google,Baidu:iru3
CircleLoss	99.57	83.42	94.00	95.73	88.75	Google,Baidu:mj00
NPCFace	99.55	83.80	94.13	95.87	89.13	Google,Baidu:2hih
MagFace	99.53	84.32	94.03	95.82	89.85	Google,Baidu:2hih

3.3 Shallow Face Learning

Training Mode	LFW	CPLFW	CALFW	AgeDb	Models&Logs
Convention Training	91.77	61.56	76.52	73.90	Google,Baidu:j4ve
Semi-siamese Training	99.38	82.53	91.78	93.60	Google,Baidu:n630

3.4 Masked Face Recognition

Model	Rank1	Rank3	Rank5	Rank10	Models&Logs	Note
model1	27.03	34.90	38.45	43.22	Google,Baidu:vp7e	Trained by MS-Celeb-1M-v1c
model2	71.40	76.60	78.62	81.05	Google,Baidu:b7tk	Trained by the upper half face in MS-Celeb-1M-v1c
model3	78.45	83.20	84.89	86.92	Google,Baidu:pcio	Trained by MS-Celeb-1M-v1c-Mask
model4	79.20	83.67	85.28	87.24	Google,Baidu:d9ii	Concat the features of model2 and model3

4. Extend the training module

4.1 Add new backbone

• Define the network under the directory backbone.
• Create the object in backbone_def.py
• Add the configuration in backbone_conf.yaml.

4.2 Add new head

• Define the new head under the directory head.
• Create the object in head_def.py
• Add the configuration in head_conf.yaml.

4.3 Add new data sampler

• Add the new data sampler in train_dataset.py

4.4 Add new training mode

• Create a new folder named by the new training mode in this directory.
• Implement the training procedure in 'train.py'.
• Add the configuration in 'train.sh'.