Time to arrival

[yonishafir]

It seems to me that you use positional encoding, although in the paper you mention to use time to arrival.

Could you please forward to where you implement and use time to arrival positional encoding?

[EricGuo5513]

https://github.com/EricGuo5513/text-to-motion/blob/main/networks/trainers.py#:~:text=%2D1%5D)-,tta%20%3D%20m_lens%20//%20self.opt.unit_length%20%2D%20i,-if%20self.

Hi, please check here.

[yonishafir]

I could not find the line where your'e using time to arrival in the above reference.

instead I see both in: https://github.com/EricGuo5513/text-to-motion/blob/main/networks/modules.py#L43

and in: https://github.com/EricGuo5513/text-to-motion/blob/main/networks/modules.py#L62

Positional encoding (forward).

I'll be happy for a better reference.

thanks

[EricGuo5513]

Hi, the time-to-arrival is technically implemented by positional encoding. While the difference is that, usually people would feed the current time step t for positional encoding, our time-to-arrival feed T-t for positional encoding, where T is the target length and t is the current time step. This could make the model know how long the generation process is yet to be done.

[yonishafir]

Thanks for the detailed answer, I couldn't find where you feed your model positional encoding with t=T-t (instead of t) can you please kindly direct me to the correct line in the trainer?

[EricGuo5513]

It is at the line 334 in https://github.com/EricGuo5513/text-to-motion/blob/main/networks/trainers.py.

[yonishafir]

Thanks a lot