INT16 model

[ShiLiangfan11]

I am currently working on video compression research and have been studying your excellent work on DCVC-RT. I noticed that your paper mentions support for INT16 inference, which is of great interest to my project. However, I could not find the corresponding testing code for INT16 in the official repository.

Would it be possible for you to share the INT16 testing code or provide guidance on how to perform INT16 inference with your model?

I would greatly appreciate any help or pointers you can offer. Thank you very much for your valuable work and contributions to the field.

[a-runner]

Have you got the INT16 testing code?

[vovasilenko]

@ShiLiangfan11, Do you have weights in the INT16 or INT8 format?

[QLaHPD]

@ShiLiangfan11, Do you have weights in the INT16 or INT8 format?

I suspect you don't need the weights in INT format, but in real time convert from fp16 to it, using the algorithm described in the paper. Right now I'm trying to implementing it, and will come back here if I succeed.

[vovasilenko]

@QLaHPD, Thank you very much for your reply, I wish you good luck! I really look forward to your success!

[QLaHPD]

@QLaHPD, Thank you very much for your reply, I wish you good luck! I really look forward to your success!

Hello, I tried to implement it without success. I'm not sure if I made a mistake (probably) or if there is some important code missing that they didn't publish. In any case, I sent an email to the authors asking for help. If you want, we can collaborate to try to implement it.

[vovasilenko]

@QLaHPD, Hello, what do you think about Triton? Is it possible to implement "integerization" for complete codec determinism?

[QLaHPD]

@QLaHPD, Hello, what do you think about Triton? Is it possible to implement "integerization" for complete codec determinism?

Hi, yes but is not worth using it in integer mode as stated in the paper, probably you would like to design a custom accelerator for that, because in int16 like in the paper, the performance drops so much that its better to use a CPU based codec like AV1.

[Joanna-0421]

@QLaHPD Hi, have you successfully implement inference with int16 format?