dog-qiuqiu / FastestDet

:zap: A newly designed ultra lightweight anchor free target detection algorithm, weight only 250K parameters, reduces the time consumption by 10% compared with yolo-fastest, and the post-processing is simpler
BSD 3-Clause "New" or "Revised" License
776 stars 133 forks source link

Fixcommit #25 #49

Open touchinglie opened 1 month ago

touchinglie commented 1 month ago

随着多个依赖的更新修复多处兼容性、多平台问题: 1、pip 不再拥有原requirements.txt内的onnxruntime==1.11.1,修改为1.16.0, 更高的版本在Windows上onnxruntime官方已经确认有问题:[1.16.2] ONNX not built correctly for Windows

2、在train.py修改保存模型名以修复Windows平台不支持冒号作为文件名的问题,修复后可以close的关联Issue有: AssertionError: 请指定正确的配置文件路径 README weight_AP05:0.253207_280-epoch.pth Colon Issues train.py导出路径 pth文件命名带了冒号 请问大佬这训练过程是否有出错的地方,一直找不到训练出来的模型 无法保存模型权重

3、onnx示例添加防止传入后预测框丢失的情况,不进行判断没有目标时进行nms会报numpy数组维度错误: IndexError: too many indices for array: array is 1-dimensional, but 2 were indexed.

4、onnx示例添加防止推理要绘制的内容带numpy的只读属性的图像矩阵复制,会报不可修改: img marked as output argument, but provided NumPy array marked as readonly

 [Opencv does not check for numpy writeable flags when writing on array.](https://github.com/opencv/opencv-python/issues/859)

opencv官方库里cv::rectangle、cv::circle(c++的,对应到Python里就是cv2.rectangle和cv2.circle)没修复这个问题的Issue: Document OpenCV 4.9 Python changes behavior with readonly numpy arrays

 修复后识别框的坐标会转换成int8整型丢失精度。