Yolov5

DeepVAC-compliant Yolov5 implementation

简介

本项目实现了符合DeepVAC规范的Yolov5

项目依赖

• deepvac >= 0.6.0
• pytorch >= 1.9.0
• torchvision >= 0.10.0

如何运行本项目

1. 阅读DeepVAC规范

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

3. 准备数据集

• 获取coco2017数据集
  浏览器操作：
  coco2017labels.zip
  train2017.zip
  val2017.zip
  test2017.zip

命令行操作：

curl -L https://github.com/ultralytics/yolov5/releases/download/v1.0/coco2017labels.zip -o coco2017labels.zip
mkdir -p data/coco
unzip -q coco2017labels -d data/coco   
rm coco2017labels.zip

• 解压coco2017数据集

• 数据集配置 在config.py文件中作如下配置：

from deepvac import AttrDict, new

# new("your train class name")
config = new("Yolov5Train")

# train dataset
train_sample_path = "data/coco/images/train2017"
train_target_path = "data/coco/instances_train2017.json"
config.core.Yolov5Train.train_dataset = Yolov5MosaicDataset(config, train_sample_path, train_target_path, config.core.Yolov5Train.img_size, config.core.Yolov5Train.border)

# val dataset
val_sample_path = "data/coco/images/val2017"
val_target_path = "data/coco/instances_val2017.json"
config.core.Yolov5Train.val_dataset = Yolov5Dataset(config, val_sample_path, val_target_path, config.core.Yolov5Train.img_size)

# test dataset
config.core.Yolov5Test = AttrDict()
config.core.Yolov5Test.test_sample_path = "your test images dir"

• 如果是自己的数据集，那么必须要符合标准coco标注格式

4. 训练相关配置

• 指定预训练模型路径(config.core.Yolov5Train.model_path)
  yolov5s & yolov5l 提取码: g4tu
• 指定训练分类数量(config.core.Yolov5Train.class_num)
• 是否采用混合精度训练(config.core.Yolov5Train.amp)
• 是否采用ema策略(config.core.Yolov5Train.ema)
• 是否采用梯度积攒到一定数量在进行反向更新梯度策略(config.core.Yolov5Train.nominal_batch_factor)
• dataloader相关配置(config.core.Yolov5Train.num_workers)

config.core.Yolov5Train.model_path = "output/pretrained.pth"

config.core.Yolov5Train.class_num = 80

config.core.Yolov5Train.amp = False

config.core.Yolov5Train.ema = True
# define ema_decay with other func
# config.ema_decay = lambda x: 0.9999 * (1 - math.exp(-x / 2000))

config.core.Yolov5Train.nominal_batch_factor = 4

config.core.Yolov5Train.shuffle = True
config.core.Yolov5Train.batch_size = 16
config.core.Yolov5Train.num_workers = 8
config.core.Yolov5Train.pin_memory = True

5. 训练

5.1 单卡训练

执行命令：

python3 train.py

6. 测试

• 测试相关配置

config.core.Yolov5Test.device = "cuda"
config.core.Yolov5Test.class_num = 80
config.core.Yolov5Test.img_size = 640
config.core.Yolov5Test.half = False
config.core.Yolov5Test.show_output_dir = "output/show"
config.core.Yolov5Test.iou_thres = 0.45
config.core.Yolov5Test.conf_thres = 0.25
config.core.Yolov5Test.idx2cat = ["cls{}".format(i) for i in range(config.core.Yolov5Test.class_num)]

• 运行测试脚本：

# 方法1
config.core.Yolov5Test.model_path = <trained-model>
config.core.Yolov5Test.test_sample_path = <test-sample-path>
python3 test.py

# 方法2
python3 test.py <trained-model(required)> <test-sample-path(required)> <label-path(optional)>

7. 使用torchscript模型

如果训练过程中未开启config.core.Yolov5Test.script_model_path开关，可以在测试过程中转化torchscript模型

• 转换torchscript模型(*.pt)

config.core.Yolov5Test.ema = False
config.core.Yolov5Test.script_model_path = "output/script.pt"

按照步骤6完成测试，torchscript模型将保存至config.core.Yolov5Test.script_model_path指定文件位置

• 加载torchscript模型

config.core.Yolov5Test.jit_model_path = <torchscript-model-path>

8. 使用静态量化模型

TODO

9. 使用coreml模型

• 测试环境

torch == 1.8.1
numpy == 1.19.5
coremltools == 4.1

• 转换coreml模型(*.mlmodel)

# 如果指定路径，那么在训练或者测试过程中都会进行模型转换
import coremltools
config.cast.CoremlCast = AttrDict()
config.cast.TraceCast = AttrDict()
config.cast.TraceCast.model_dir = "output/trace.pt"
config.cast.CoremlCast.model_dir = "output/coreml.mlmodel"
config.cast.CoremlCast.input_type = None
config.cast.CoremlCast.scale = 1.0 / 255.0
config.cast.CoremlCast.color_layout = 'BGR'
config.cast.CoremlCast.blue_bias = 0
config.cast.CoremlCast.green_bias = 0
config.cast.CoremlCast.red_bias = 0
config.cast.CoremlCast.minimum_deployment_target = coremltools.target.iOS13
config.cast.CoremlCast.classfier_config = ["cls{}".format(i) for i in range(config.core.Yolov5Test.class_num)]

# 训练过程中转换(cast2cpu配置可配置在config.py文件也可以配置在train.py文件)
config.core.Yolov5Train.cast2cpu = True
python3 train.py
# 测试过程中转换(cast2cpu配置可配置在config.py文件也可以配置在train.py文件)
config.core.Yolov5Test.cast2cpu = True
python3 test.py <trained-model(required)> <test-sample-path(required)> <label-path(optional)>

• coreml模型推理
  推理环境: macos13 or later

10. 更多功能

如果要在本项目中开启如下功能：

• 使用tensorboard
• 转换ONNX
• 转换NCNN
• 开启量化
• 开启自动混合精度训练

请参考DeepVAC

11. TODO

• 20210201 项目增加了对Yolov5S和Yolov5L的支持
• 20210219 修复了torchscript模型C++推理代码在cuda上CUDNN_STATUS_INTER_ERROR问题(在modules/model.py中重写Fcous模块）
• 修复在test过程中，静态量化模型报错问题
• 20210622 项目更新来适应新版本deepvac
• 20210625 项目增加coreml模型转换