Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/deepvac/retinaface
DeepVAC-compliant RetinaFace implementation
https://github.com/deepvac/retinaface
Last synced: about 2 months ago
JSON representation
DeepVAC-compliant RetinaFace implementation
- Host: GitHub
- URL: https://github.com/deepvac/retinaface
- Owner: DeepVAC
- License: gpl-3.0
- Created: 2021-01-26T11:08:54.000Z (almost 4 years ago)
- Default Branch: main
- Last Pushed: 2021-06-22T15:31:13.000Z (over 3 years ago)
- Last Synced: 2023-03-05T16:49:50.373Z (almost 2 years ago)
- Language: Python
- Size: 115 KB
- Stars: 1
- Watchers: 2
- Forks: 1
- Open Issues: 1
-
Metadata Files:
- Readme: README.md
- License: LICENSE
Awesome Lists containing this project
README
# RetinaFace
DeepVAC-compliant RetinaFace implementation
# 简介
本项目实现了符合DeepVAC规范的RetinaFace 。
### 项目依赖
- deepvac >= 0.5.7
- pytorch >= 1.8.0
- torchvision >= 0.7.0
- opencv-python
- numpy
# 如何运行本项目
## 1. 阅读[DeepVAC规范](https://github.com/DeepVAC/deepvac)
可以粗略阅读，建立起第一印象。
## 2. 准备运行环境
使用Deepvac规范指定[Docker镜像](https://github.com/DeepVAC/deepvac#2-%E7%8E%AF%E5%A2%83%E5%87%86%E5%A4%87)。
## 3. 准备数据集
- 获取WIDER Face数据集
[WIDER Face Training Images](https://share.weiyun.com/5WjCBWV)
[WIDER Face Testing Images](https://share.weiyun.com/5vSUomP)
[Face annotations](http://mmlab.ie.cuhk.edu.hk/projects/WIDERFace/support/example/Submission_example.zip)
- 若想进一步了解WIDER Face数据集，可参考官网上的信息。
[WIDER Face官网](http://shuoyang1213.me/WIDERFACE)
- 解压WIDER Face数据集
- 数据集配置
在config.py文件中作如下配置：
```python
# line 26
config.datasets.RetinaTrainDataset.fileline_path =
# line 27
config.datasets.RetinaTrainDataset.sample_path_prefix =
# line 76
config.sample_path =
```
- 如果是自己的数据集，那么必须要跟widerface的标注格式一致
## 4. 训练相关配置
- 指定预训练模型路径(config.core.RetinaTrain.model_path)
- 指定Backbone网络结构, 支持ResNet50, MobileNetV3, RegNet, RepVGG(config.core.RetinaTrain.net)
- 指定loss函数(config.core.RetinaTrain.criterion)
- 指定训练分类数量(config.core.RetinaTrain.class_num)
- 指定优化器optimizer(config.core.RetinaTrain.optimizer)
- 指定学习率策略scheduler(config.core.RetinaTrain.scheduler)
```python
config.core.RetinaTrain.model_path = ''
config.core.RetinaTrain.class_num = 2
config.core.RetinaTrain.shuffle = True
config.core.RetinaTrain.batch_size = 24
config.core.RetinaTrain.net = RetinaFaceMobileNet()
config.core.RetinaTrain.criterion = MultiBoxLoss(config.train.cls_num, 0.35, True, 0, True, 7, 0.35, False, config.train.device)
config.core.RetinaTrain.optimizer = torch.optim.SGD(
config.core.RetinaTrain.net.parameters(),
lr=1e-3,
momentum=0.9,
weight_decay=5e-4,
nesterov=False
)
config.core.RetinaTrain.scheduler = optim.lr_scheduler.MultiStepLR(config.core.RetinaTrain.optimizer, [100, 150, 190, 220], 0.1)
```
## 5. 训练
### 5.1 单卡训练
执行命令：
```bash
python3 train.py
```
### 5.2 分布式训练
在config.py中修改如下配置：
```python
#dist_url，单机多卡无需改动，多机训练一定要修改
config.core.RetinaTrain.dist_url = "tcp://localhost:27030"
#rank的数量，一定要修改
config.core.RetinaTrain.world_size = 2
```
然后执行命令：
```bash
python train.py --rank 0 --gpu 0
python train.py --rank 1 --gpu 1
```
## 6. 测试
- 测试相关配置
```python
# config.core.RetinaTest is config used for post_process and retina_test.
config.core.RetinaTest.model_path = ""
config.core.RetinaTest.confidence_threshold = 0.02
config.core.RetinaTest.nms_threshold = 0.4
config.core.RetinaTest.top_k = 5000
config.core.RetinaTest.keep_top_k = 1
# config.core.FaceTest is config used for face end-to-end test.
# align type
config.core.FaceTest.align_type = ['align', 'no_align', 'warp_crop']
# db/ds path and prefix(name)
config.core.FaceTest.test_dirs = ['']
config.core.FaceTest.test_prefix = ['']
config.core.FaceTest.db_dirs = ['']
config.core.FaceTest.db_prefix = ['']
# config.core.FaceRecTest is config used in face recognition module.
config.core.FaceRecTest.jit_model_path = ""
```
- 加载模型(*.pth)
```python
config.core.RetinaTest.model_path =
```
- 运行测试脚本：
```bash
python3 test.py
```
## 7. 使用trace模型/script模型
如果训练过程中开启config.cast.TraceCast（或者config.cast.ScriptCast)开关，可以在测试过程中转化torchscript模型
- 转换torchscript模型(*.pt)
```python
# trace
config.cast.TraceCast = AttrDict()
config.cast.TraceCast.model_dir = "./trace.pt"
# script
config.cast.ScriptCast = AttrDict()
config.cast.ScriptCast.model_dir = "./script.pt"
```
按照步骤6完成测试，torchscript模型将保存至model_dir指定文件位置
- 加载torchscript模型
```python
config.core.RetinaTest.jit_model_path =
```
## 8. 使用静态量化模型
如果训练过程中未开启config.cast.TraceCast开关，可以在测试过程中转化静态量化模型
- 转换静态模型(*.sq)
```python
# trace
config.cast.TraceCast.static_quantize_dir = "./trace.sq"
# script
config.cast.ScriptCast.static_quantize_dir = "./script.sq"
```
按照步骤6完成测试，静态量化模型将保存至config.static_quantize_dir指定文件位置
- 加载静态量化模型
```python
config.core.RetinaTest.jit_model_path =
```
- 动态量化模型对应的配置参数为config.cast.TraceCast.dynamic_quantize_dir(或者config.cast.ScriptCast.dynamic_quantize_dir)
## 9. 更多功能
如果要在本项目中开启如下功能：
- 预训练模型加载
- checkpoint加载
- 使用tensorboard
- 启用TorchScript
- 转换ONNX
- 转换NCNN
- 转换CoreML
- 开启量化
- 开启自动混合精度训练
请参考[DeepVAC](https://github.com/DeepVAC/deepvac)