Jiaming Han

建议重新下载checkpoint

参考我们的：https://github.com/csuhan/ReDet/blob/master/GETTING_STARTED.md

我们目前不支持online evaluation，因此需要将`valset`当作`test`，单独运行train->test(val)->evaluation.

We have tested on Pytorch1.1/1.3. Try lower version Pytorch or see [here](https://github.com/sgrvinod/a-PyTorch-Tutorial-to-Object-Detection/issues/24)

@why228430 Have you re-compiled the ops after your re-installation of Pytorch? Please del all *.so and re-compile the ops.

@1127244933 What is your GPU? I have tested on V100 and Titan Xp with Pytorch=1.3.1, cuda=10.1. Besides, is the inference (demo_large_image.py) ok?

For the `polyiou`, you need: ``` sudo apt-get install swig cd DOTA_devkit swig -c++ -python polyiou.i python setup.py build_ext --inplace ``` For the `riroi_align`, I suggest to change `THREADS_PER_BLOCK` from...

As I am not familiar with CUDA programming, the current code is not so efficient and robust. I will continue to optimize the code, and pulling requests are welcomed!

建议1000*1000，mask会相应变大

请将你的数据转换类似于DOTA的格式，然后再按照repo中介绍的pipeline。