David Zhu

Multi-GPU VS

1

Hi, how to make gnina perform multi-GPU parallel computing in virtual screening tasks? The current speed (e.g. --cnn_scoring rescore) on a single GPU is not satisfactory for, ...maybe most of...

Running speed

2

How fast is GNINA? Is there any comparison with the Vina-GPU version?

automatic processing tools for map.fld file

1

Hi, is there any tool for automatic processing of map.fld file generated from receptor? If I want to prepare multiple proteins, such as proteins from pdbbind, manual processing is very...

Rotated IOU Loss Training showed NaN in custom dataset

2

Hi！请问用rotated IOU loss是不是容易引起梯度爆炸？我在自定义的数据上表现出的梯度爆炸的情况，刚开始是没问题的，就是loss有点抖，但是后面就直接nan了；IOU loss应该不可导吧，请问这个Rotated IOU loss是怎么实现可微的，我看上交的严老师课题组基于Ratated IOU loss不可导的问题发展了两套近似的loss，比较好奇你们怎么做到可导的。期待您的答复~

Anchor generator

1

Hi!从你的源码中看出，FAM模块内似乎要求每个grid的anchor只能存在一种预先设定好的情况，因为这样才允许后续的Feature alignment。在交叉的立交桥等存在object交叉重叠的场景下，上s2anet检测的精度下降了，只能检测交叉重叠物体的其中一个。虽然ODM模块是可以引入多个anchors的，但是因为存在FAM和ODM，输入到ODM模块的可能也是交叉物体中的其中一个，然后通过rotation的最大响应，提取其中的一个交叉物体，我认为s2anet提升交叉重叠物体的检测精度需要从增加FAM模块anchor入手。请问有什么解决方案么？

When I generated tmap following the drugBank example in your document, TMap generated HTML files normally, but did not display tmap tree.

cmake error

1

When I `cmake ..`, error found as below: CMake Error at /opt/conda/lib/cmake/Boost-1.70.0/BoostConfig.cmake:95 (find_package): Could not find a package configuration file provided by "boost_python38" (requested version 1.70.0) with any of the...

GD Loss Training showed NaN in custom dataset

8

Hi！请问用GD loss（KLD和GWD）是不是容易引起梯度爆炸？我在自定义的数据上表现出的梯度爆炸的情况，前几个iteration还ok，但是后面就直接nan了，期待您的答复~