PSENet

DeepVAC-compliant PSENet implementation.

简介

本项目实现了符合DeepVAC规范的OCR检测模型PSENet

项目依赖

• deepvac >= 0.5.6
• pytorch >= 1.8.0
• torchvision >= 0.7.0
• opencv-python
• numpy
• pyclipper (pip install pyclipper)
• Polygon3 (pip install Polygon3)
• pillow

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

• 获取文本检测数据集 CTW1500格式的数据集,CTW1500下载地址:

  ch4_training_images.zip

  ch4_training_localization_transcription_gt.zip

  ch4_test_images.zip

  Challenge4_Test_Task1_GT.zip

• 数据集配置 在config.py文件中作如下配置：

sample_path = <your train/val/test image path>
label_path = <your train/val gt path>

4. 训练相关配置

• dataloader相关配置

is_transform = True                 # 是否动态数据增强
img_size = 640                      # 输入图片大小(img_size, img_size)
config.datasets.PseTrainDataset = AttrDict()
config.datasets.PseTrainDataset.kernel_num = 7
config.datasets.PseTrainDataset.min_scale = 0.4

config.core.PSENetTrain.batch_size = 12
config.core.PSENetTrain.num_workers = 4
config.core.PSENetTrain.train_dataset = PseTrainDataset(config, sample_path, label_path, is_transform, img_size)
config.core.PSENetTrain.train_loader = torch.utils.data.DataLoader(
    dataset = config.core.PSENetTrain.train_dataset,
    batch_size = config.core.PSENetTrain.batch_size,
    shuffle = True,
    num_workers = config.core.PSENetTrain.num_workers,
    pin_memory = True,
)

5. 训练

python3 train.py

6. 测试

• 编译PSE

cd modules/cpp
make

• 测试相关配置

config.core.PSENetTest = config.core.PSENetTrain.clone()
config.core.PSENetTest.model_path = "your test model dir / pretrained weights"
config.core.PSENetTest.kernel_num = 7
config.core.PSENetTest.min_kernel_area = 10.0
config.core.PSENetTest.min_area = 300.0
config.core.PSENetTest.min_score = 0.93
config.core.PSENetTest.binary_th = 1.0
config.core.PSENetTest.scale = 1

sample_path = 'your test image dir'
config.core.PSENetTest.test_dataset = PseTestDataset(config, sample_path, long_size=1280)
config.core.PSENetTest.test_loader = torch.utils.data.DataLoader(
    dataset = config.core.PSENetTest.test_dataset,
    batch_size = 1,
    shuffle = False,
    num_workers = 0,
    pin_memory = True
)

• 运行测试脚本：

python3 test.py

7. 使用torchscript模型

如果训练过程中未开启config.cast.script_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.script_model_path = "output/script.pt"

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

• 加载torchscript模型

config.core.PSENetTrain.jit_model_path = <torchscript-model-path>
config.core.PSENetTest.jit_model_path = <torchscript-model-path>

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
• 采用ema策略(config.ema)
• 采用梯度积攒到一定数量在进行反向更新梯度策略(config.nominal_batch_factor)

请参考DeepVAC