ABlueLight

Any advice on how to avoid this bug

anyone can offer help? Thanks

https://github.com/zeromq/libzmq/issues/4229 https://github.com/zeromq/libzmq/issues/3937 seem is the same reason

bug is still exist

尝试使用torch==1.5.1 torchvision==0.6.1试试呢

this repo use old pytorch version. not test on newer pytorch version. maybe it not work. now i haven't much time to maintain this repo. sorry, maye use other repo...

sorry, i think there is still some problem that don't soloved.

没有在cityscapes上测试过，从头训练过coco数据，精度比官方的精度要低； 原因：1. 没有多卡训练的支持，batch设置的比较小，2. 多尺度训练的scale不够丰富，3. 可能某些细节也有待优化。 训练还是推荐使用官方的仓库。

缺陷还挺多的，现在工作比较忙，没时间维护了T_T 1. 只支持resnet18、resnet34两种backbone的训练和测试. 2. 不支持多显卡并行训练 3. 训练和测试的配置项不完整 4. eval时没有可视化 5. 导出ONNX的脚本，后续部署流程 shortages: 1. Only supports Resnet18, Resnet34 backbone to train. 2. Multi GPU parallel training is not supported. 3. Incomplete...

@seekerwmy 最好参照官方代码做一个对比:smile: