DeepVAC-compliant DBNet implementation.
本项目实现了符合DeepVAC规范的OCR检测模型DBNet
项目依赖
- deepvac >= 0.5.6
- pytorch >= 1.8.0
- torchvision >= 0.7.0
- opencv-python
- numpy
- pyclipper
- shapely
- pillow
1. 阅读DeepVAC规范
可以粗略阅读,建立起第一印象
可以使用DeepVAC规范指定的Docker镜像
-
获取文本检测数据集 CTW1500格式的数据集,CTW1500下载地址:
-
数据集配置 在config.py文件中作如下配置:
config.sample_path = <your train image path>
config.label_path = <your train gt path>
config.sample_path = <your val image path>
config.label_path = <your val gt path>
- DB backbone配置
# 目前支持resnet18,mv3large
config.arch = "resnet18"
- dataloader相关配置
config.is_transform = True # 是否做数据增强
config.img_size = 640 # 训练图片大小(img_size, img_size)
config.datasets.DBTrainDataset = AttrDict()
config.datasets.DBTrainDataset.shrink_ratio = 0.4
config.datasets.DBTrainDataset.thresh_min = 0.3
config.datasets.DBTrainDataset.thresh_max = 0.7
config.core.DBNetTrain.batch_size = 8
config.core.DBNetTrain.num_workers = 4
config.core.DBNetTrain.train_dataset = DBTrainDataset(config, config.sample_path, config.label_path, config.is_transform, config.img_size)
config.core.DBNetTrain.train_loader = torch.utils.data.DataLoader(
dataset = config.core.DBNetTrain.train_dataset,
batch_size = config.core.DBNetTrain.batch_size,
shuffle = True,
num_workers = config.core.DBNetTrain.num_workers,
pin_memory = True,
sampler = None
)
python3 train.py
- 测试相关配置
config.core.DBNetTest.model_path = <your model path> # 加载模型路径
# config.core.DBNetTest.jit_model_path = <torchscript-model-path> # torchscript model path
config.core.DBNetTest.is_output_polygon = True # 输出是否为多边形模型
config.sample_path = <your test image path> # 测试图片路径
config.core.DBNetTrain.batch_size = 8
config.core.DBNetTrain.num_workers = 4
config.core.DBNetTest.test_dataset = DBTestDataset(config, config.sample_path, long_size = 1280)
config.core.DBNetTest.test_loader = torch.utils.data.DataLoader(
dataset = config.core.DBNetTest.test_dataset,
batch_size = config.core.DBNetTrain.batch_size,
shuffle = False,
num_workers = config.core.DBNetTrain.num_workers,
pin_memory = True
)
- 运行测试脚本:
python3 test.py
如果训练过程中未开启config.cast.TraceCast.model_dir开关,可以在测试过程中转化torchscript模型
- 转换torchscript模型(.pt)
config.cast.TraceCast.model_dir = "output/script.pt"
按照步骤6完成测试,torchscript模型会保存至config.cast.TraceCast.model_dir指定位置
- 加载torchscript模型
config.core.DBNetTest.jit_model_path = <torchscript-model-path>
然后按照步骤6测试,会读取script_model
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
- 采用ema策略(config.ema)
- 采用梯度积攒到一定数量再进行反向更新梯度策略(config.nominal_batch_factor)
请参考DeepVAC