SHL

你自己数据集样本太少了，得先加载预训练的backbone才行


ap50: 50%重合度阈值下的准确度（工业界更看重这个） ap50:95：50%、55%、...、95%重合度阈值下的准确度的平均数（学术上更看重这个）

按道理是不需要修改别的地方的，方便把模型发给我看看吗 [email protected]

方便加微信聊吗，微信号已发您邮箱

just set num_workers=4. Too large number of num_workers may cause the opposite effect.

https://github.com/NVIDIA/TensorRT/issues/1634


in edgeyolo/export/calib.py, change this function to ```python class CalibDataset(Dataset): def __getitem__(self, item): import numpy as np ret = [] for file in self.path_list[item:item + self.batch]: im, _ = preproc(cv2.imread(file), self.input_size)...

I don't know whether it is useful because i didn't occur this problem. You can have a try