SHL
本工作的主要提升点在AP50上，速度方面都大差不差，不占优势。 这是两个多月前做的工作了，只不过前两天才挂到arxiv上，如果真要用的话还是建议yolov6 v3.0版本，肯定更好。 可以试试把yolov6的检测头魔改成我们的，应该会更快些。
在投^_^
[edgeyolo/data/datasets/enhanced_mosaicdetection.py](https://github.com/LSH9832/edgeyolo/blob/main/edgeyolo/data/datasets/enhanced_mosaicdetection.py). Some files such as [edgeyolo/utils2/datasets.py](https://github.com/LSH9832/edgeyolo/blob/main/edgeyolo/utils2/datasets.py) in this repository are not used actually.
try set "cache: False" in your dataset yaml
这是很早之前下的代码吧，我记得已经修复了，你重新下一遍
just modify the last line in your config file params/train/train_XXX.yaml as follows ``` force_start_epoch: 0 # set -1 to disable this option ```
Not yet, but YOLO format is similar to Visdrone formats(both load labels from txt), and its dataloader locates in path [edgeyolo/data/datasets/visdrone.py](https://github.com/LSH9832/edgeyolo/blob/main/edgeyolo/data/datasets/visdrone.py), I think just modify the following function and then...
参数量已在表格中。计算量（GFLOPS）在不同系统不同硬件上会有略微差异，所以没写，建议到自己的设备上测。
see params/train/train_settings.yaml ``` input_size: [672, 672] # image input size for model multiscale_range: 5 # real_input_size = input_size + randint(-multiscale_range, multiscale_range) * 32 ``` so if you want to use...
some device do not support fp16 precision, in this case do not use option "--fp16"