search

HKUST-Aerial-Robotics / EPSILON

MIT License
564 stars 205 forks source link

How to access the neural network module? #18

Closed cugbbaiyun closed 2 years ago

cugbbaiyun commented 2 years ago

Hi, I read the code but i can hardly find the definitions of neural network modules (encoder and decoder in intention estimator) and the parameters. It seems that the neural network modules are not included in encoder.h and decoder.h. And is it possible to change the structure of encoders or decoders? Besides, how to train my own decoders and encoders?

MasterIzumi commented 2 years ago

Unfortunately, the learning-based intention estimation part is not available in the released repo. Actually, it is a pretty simple network compared to the current prediction SOTAs (of course there's a quite big performance gap as well). If you really need this part, we can provide you with a sample code only for the learning-based intention prediction module. (Have responded via email)

haotruongnhat commented 1 year ago

Hi authors, could I have the code of learning-based intention prediction module for reference. Thank a lot! Regards,

feiyuxiaoThu commented 4 months ago

Unfortunately, the learning-based intention estimation part is not available in the released repo. Actually, it is a pretty simple network compared to the current prediction SOTAs (of course there's a quite big performance gap as well). If you really need this part, we can provide you with a sample code only for the learning-based intention prediction module. (Have responded via email)

We recently want to evaluate the performance of the distilled-Bi-LSTM network (from some transformer network), and could you provide your basic network so we can use for baseline. Thanks a lot for your reply.

MasterIzumi commented 4 months ago

@feiyuxiaoThu Actually, it is a quite simple intention classification network and already out-of-date at this time point. If you are working on the motion or trajectory prediction task, you can refer to our recent work which is more modern and effective. Project page: https://github.com/HKUST-Aerial-Robotics/SIMPL

feiyuxiaoThu commented 4 months ago

@feiyuxiaoThu Actually, it is a quite simple intention classification network and already out-of-date at this time point. If you are working on the motion or trajectory prediction task, you can refer to our recent work which is more modern and effective. Project page: https://github.com/HKUST-Aerial-Robotics/SIMPL

Thanks for your reply, and we will refer to your mentioned work.

wlsdaydayup commented 1 month ago

Hi authors, could I have the code of learning-based intention prediction module of EPSILON for reference. Thank a lot! Regards, Myemail:sduwls@163.com