tanjatang / dcenet Goto Github PK

Hello, it seems difficult to understand the calculation of the KL_loss for me. Can someone give me a few tips or references?

Hi,

I like your take on ranking the trajectories. I believe it is important and have not seen anyone try to solve it before. That said, I am not 100% sure if I have understood it properly and hence few questions.

Question 1

In the paper, you state that you follow graves (https://arxiv.org/pdf/1308.0850.pdf), to come up with the PDF

I then read the referenced paper of Alex Graves and to me, it seems that he is suggesting to use Mixture Density Networks to learn the distribution. Based on that I would have thought there will be an MDN layer present after the decoder.

In the source code, it seems that you take the empirical mean of the predicted positions for an agent, and then that is used to create the PDF.

Am sure I am missing something fundamental here. Would appreciate it if you could educate.

Question 2

In your implementation, before you rank the trajectories, you bring the predicted trajectories to the original coordinate system by adding the last observation.

I was wondering if there is any downside to ranking the trajectories before bringing it them original coordinate. The reason I am thinking in this direction is that if there is no downside then it could save some computation time. For e.g. you could rank the trajectory and find the most likely one and then bring only this most likely one to the original coordinate system. This way addition of the last observation to the normalized trajectory is to be done on 1 trajectory per agent instead of 25 trajectories per agent. The same applies to the cumsum operation etc

Thanks in advance.

Regards & thanks
Kapil

Hello,
when applying the pretrained model best.hdf5, the test performance of bookstore-3 is shown as below:

it doesn't look like the one in the paper, all the prediction lines are just straight, I am wondering about it, could you please give me some advice on how it's happening? Thx and wish for you reply.

Hi,

First, thank you for the nicely written paper and for publishing the source code as well.

I was browsing through your code and noticed this -

Based on my understanding, because of this hardcode training=False, it seems that dropouts in TransformerBlock layer will not happen even during training.

May be intentional but just wanted to point it out in case it is a typo

Regards & thanks
Kapil

Hello @tanjatang @haohao11,

thank you for your work!

I have some questions if you can help me please:
1- I need details on generating trajectories for multiple agents.
2- How are the interactions between all predictions handled?
3- How is the nature of the agents taken into account in the generation of trajectories / interactions? (categorization of different classes of agents..)

Thanks