Skip to content

Commit a5d5b4f

Browse files
anxiangsiranxiangsir
committed
updated README.md
1 parent cd7a669 commit a5d5b4f

File tree

1 file changed

+56
-20
lines changed

1 file changed

+56
-20
lines changed

‎recognition/arcface_torch/README.md‎

Lines changed: 56 additions & 20 deletions
Original file line numberDiff line numberDiff line change
@@ -5,9 +5,9 @@ identity on a single server.
55

66
## Requirements
77

8-
- Install [PyTorch](http://pytorch.org) (torch>=1.6.0), our doc for [install.md](docs/install.md).
8+
- Install [PyTorch](http://pytorch.org) (torch>=1.9.0), our doc for [install.md](docs/install.md).
99
- (Optional) Install [DALI](https://docs.nvidia.com/deeplearning/dali/user-guide/docs/), our doc for [install_dali.md](docs/install_dali.md).
10-
- `pip install -r requirement.txt`.
10+
- `pip install -r requirements.txt`.
1111

1212
## How to Training
1313

@@ -58,26 +58,55 @@ For **ICCV2021-MFR-ALL** set, TAR is measured on all-to-all 1:1 protocal, with F
5858
globalised multi-racial testset contains 242,143 identities and 1,624,305 images.
5959

6060

61+
> 1. Large Scale Datasets
62+
63+
| Datasets | Backbone | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) | Training Throughout | log |
64+
|:-----------------|:------------|:------------|:------------|:------------|:--------------------|:------------------------------------------------------------------------------------------------------------------------------------------------|
65+
| MS1MV3 | mobileface | 65.76 | 94.44 | 91.85 | ~13000 | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_mobileface_lr02/training.log) |
66+
| Glint360K | mobileface | 69.83 | 95.17 | 92.58 | -11000 | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_mobileface_lr02_bs4k/training.log) |
67+
| WF42M-PFC-0.2 | mobileface | 73.80 | 95.40 | 92.64 | (16GPUs)~18583 | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_mobilefacenet_pfc02_bs8k_16gpus/training.log) |
68+
| MS1MV3 | r100 | 83.23 | 96.88 | 95.31 | ~3400 | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_r100_lr02/training.log) |
69+
| Glint360K | r100 | 90.86 | 97.53 | 96.43 | ~5000 | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_r100_lr02_bs4k_16gpus/training.log) |
70+
| WF42M-PFC-0.2 | r50(bs4k) | 93.83 | 97.53 | 96.16 | (8 GPUs)~5900 | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r50_bs4k_pfc02/training.log) |
71+
| WF42M-PFC-0.2 | r50(bs8k) | 93.96 | 97.46 | 96.12 | (16GPUs)~11000 | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r50_lr01_pfc02_bs8k_16gpus/training.log) |
72+
| WF42M-PFC-0.2 | r50(bs4k) | 94.04 | 97.48 | 95.94 | (32GPUs)~17000 | click me |
73+
| WF42M-PFC-0.0018 | r100(bs16k) | 93.08 | 97.51 | 95.88 | (32GPUs)~10000 | click me |
74+
| WF42M-PFC-0.2 | r100(bs4k) | 96.69 | 97.85 | 96.63 | (16GPUs)~5200 | [click me](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r100_bs4k_pfc02/training.log) |
75+
76+
> 2. VIT For Face Recognition
77+
78+
| Datasets | Backbone | FLOPs | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) | Training Throughout | log |
79+
|:--------------|:-------------|:------|:------------|:------------|:------------|:--------------------|:---------|
80+
| WF42M-PFC-0.3 | R18(bs4k) | 2.6 | 79.13 | 95.77 | 93.36 | - | click me |
81+
| WF42M-PFC-0.3 | R50(bs4k) | 6.3 | 94.03 | 97.48 | 95.94 | - | click me |
82+
| WF42M-PFC-0.3 | R100(bs4k) | 12.1 | 96.69 | 97.82 | 96.45 | - | click me |
83+
| WF42M-PFC-0.3 | R200(bs4k) | 23.5 | 97.70 | 97.97 | 96.93 | - | click me |
84+
| WF42M-PFC-0.3 | VIT-T(bs24k) | 1.5 | 92.24 | 97.31 | 95.97 | (64GPUs)~35000 | click me |
85+
| WF42M-PFC-0.3 | VIT-S(bs24k) | 5.7 | 95.87 | 97.73 | 96.57 | (64GPUs)~25000 | click me |
86+
| WF42M-PFC-0.3 | VIT-B(bs24k) | 11.4 | 97.42 | 97.90 | 97.04 | (64GPUs)~13800 | click me |
87+
| WF42M-PFC-0.3 | VIT-L(bs24k) | 25.3 | 97.85 | 98.00 | 97.23 | (64GPUs)~9406 | click me |
88+
89+
WF42M means WebFace42M, `PFC-0.3` means negivate class centers sample rate is 0.3.
90+
91+
> 3. Noisy Datasets
92+
93+
| Datasets | Backbone | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) | log |
94+
|:-------------------------|:---------|:------------|:------------|:------------|:---------|
95+
| WF12M-Flip(40%) | R50 | 43.87 | 88.35 | 80.78 | click me |
96+
| WF12M-Flip(40%)-PFC-0.3* | R50 | 80.20 | 96.11 | 93.79 | click me |
97+
| WF12M-Conflict | R50 | 79.93 | 95.30 | 91.56 | click me |
98+
| WF12M-Conflict-PFC-0.3* | R50 | 91.68 | 97.28 | 95.75 | click me |
99+
100+
WF12M means WebFace12M, `+PFC-0.3*` denotes additional abnormal inter-class filtering.
61101

62-
| Datasets | Backbone | **MFR-ALL** | IJB-C(1E-4) | IJB-C(1E-5) | Training Throughout | log |
63-
|:-------------------------|:-----------|:------------|:------------|:------------|:--------------------|:------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
64-
| MS1MV3 | mobileface | 65.76 | 94.44 | 91.85 | ~13000 | [log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_mobileface_lr02/training.log)\|[config](configs/ms1mv3_mobileface_lr02.py) |
65-
| Glint360K | mobileface | 69.83 | 95.17 | 92.58 | -11000 | [log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_mobileface_lr02_bs4k/training.log)\|[config](configs/glint360k_mobileface_lr02_bs4k.py) |
66-
| WebFace42M-PartialFC-0.2 | mobileface | 73.80 | 95.40 | 92.64 | (16GPUs)~18583 | [log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_mobilefacenet_pfc02_bs8k_16gpus/training.log)\|[config](configs/webface42m_mobilefacenet_pfc02_bs8k_16gpus.py) |
67-
| MS1MV3 | r100 | 83.23 | 96.88 | 95.31 | ~3400 | [log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/ms1mv3_r100_lr02/training.log)\|[config](configs/ms1mv3_r100_lr02.py) |
68-
| Glint360K | r100 | 90.86 | 97.53 | 96.43 | ~5000 | [log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/glint360k_r100_lr02_bs4k_16gpus/training.log)\|[config](configs/glint360k_r100_lr02_bs4k_16gpus.py) |
69-
| WebFace42M-PartialFC-0.2 | r50(bs4k) | 93.83 | 97.53 | 96.16 | (8 GPUs)~5900 | [log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r50_bs4k_pfc02/training.log)\|[config](configs/webface42m_r50_lr01_pfc02_bs4k_8gpus.py) |
70-
| WebFace42M-PartialFC-0.2 | r50(bs8k) | 93.96 | 97.46 | 96.12 | (16GPUs)~11000 | [log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r50_lr01_pfc02_bs8k_16gpus/training.log)\|[config](configs/webface42m_r50_lr01_pfc02_bs8k_16gpus.py) |
71-
| WebFace42M-PartialFC-0.2 | r50(bs4k) | 94.04 | 97.48 | 95.94 | (32GPUs)~17000 | log\|[config](configs/webface42m_r50_lr01_pfc02_bs4k_32gpus.py) |
72-
| WebFace42M-PartialFC-0.2 | r100(bs4k) | 96.69 | 97.85 | 96.63 | (16GPUs)~5200 | [log](https://raw.githubusercontent.com/anxiangsir/insightface_arcface_log/master/webface42m_r100_bs4k_pfc02/training.log)\|[config](configs/webface42m_r100_lr01_pfc02_bs4k_16gpus.py) |
73-
| WebFace42M-PartialFC-0.2 | r200 | - | - | - | - | log\|config |
74102

75-
`PartialFC-0.2` means negivate class centers sample rate is 0.2.
76103

77104

78105
## Speed Benchmark
106+
<div><img src="https://github.com/anxiangsir/insightface_arcface_log/blob/master/pfc_exp.png" width = "90%" /></div>
107+
79108

80-
`arcface_torch` can train large-scale face recognition training set efficiently and quickly. When the number of
109+
**Arcface-Torch** can train large-scale face recognition training set efficiently and quickly. When the number of
81110
classes in training sets is greater than 1 Million, partial fc sampling strategy will get same
82111
accuracy with several times faster training performance and smaller GPU memory.
83112
Partial FC is a sparse variant of the model parallel architecture for large sacle face recognition. Partial FC use a
@@ -86,12 +115,12 @@ sparse part of the parameters will be updated, which can reduce a lot of GPU mem
86115
we can scale trainset of 29 millions identities, the largest to date. Partial FC also supports multi-machine distributed
87116
training and mixed precision training.
88117

89-
![Image text](https://github.com/anxiangsir/insightface_arcface_log/blob/master/partial_fc_v2.png)
118+
90119

91120
More details see
92121
[speed_benchmark.md](docs/speed_benchmark.md) in docs.
93122

94-
### 1. Training speed of different parallel methods (samples / second), Tesla V100 32GB * 8. (Larger is better)
123+
> 1. Training speed of different parallel methods (samples / second), Tesla V100 32GB * 8. (Larger is better)
95124
96125
`-` means training failed because of gpu memory limitations.
97126

@@ -104,7 +133,7 @@ More details see
104133
| 16000000 | **-** | **-** | 2679 |
105134
| 29000000 | **-** | **-** | **1855** |
106135

107-
### 2. GPU memory cost of different parallel methods (MB per GPU), Tesla V100 32GB * 8. (Smaller is better)
136+
> 2. GPU memory cost of different parallel methods (MB per GPU), Tesla V100 32GB * 8. (Smaller is better)
108137
109138
| Number of Identities in Dataset | Data Parallel | Model Parallel | Partial FC 0.1 |
110139
|:--------------------------------|:--------------|:---------------|:---------------|
@@ -126,11 +155,18 @@ More details see
126155
pages={4690--4699},
127156
year={2019}
128157
}
158+
@inproceedings{an2022pfc,
159+
title={Killing Two Birds with One Stone: Efficient and Robust Training of Face Recognition CNNs by Partial FC},
160+
author={An, Xiang and Deng, Jiangkang and Guo, Jia and Feng, Ziyong and Zhu, Xuhan and Jing, Yang and Tongliang, Liu},
161+
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
162+
year={2022}
163+
}
129164
@inproceedings{an2020partical_fc,
130165
title={Partial FC: Training 10 Million Identities on a Single Machine},
131166
author={An, Xiang and Zhu, Xuhan and Xiao, Yang and Wu, Lan and Zhang, Ming and Gao, Yuan and Qin, Bin and
132167
Zhang, Debing and Fu Ying},
133-
booktitle={Arxiv 2010.05222},
168+
booktitle={Proceedings of International Conference on Computer Vision Workshop},
169+
pages={1445-1449},
134170
year={2020}
135171
}
136172
```

0 commit comments

Comments
 (0)