Model Size

[Garfield2005]

I noticed that Utonia uses different model sizes compared to Concerto. Why is there this change?

https://huggingface.co/Pointcept/Utonia

Model Size	Channels	Depths	Head nums
Utonia	(54, 108, 216, 432, 576)	(3, 3, 3, 12, 3)	(3, 6, 12, 24, 32)

https://huggingface.co/Pointcept/Concerto

Model Size	Channels	Depths	Head nums
Tiny	(16, 32, 64, 128, 256)	(1, 1, 1, 3, 1)	(1, 2, 4, 8, 16)
Small	(32, 64, 128, 256, 512)	(2, 2, 2, 6, 2)	(2, 4, 8, 16, 32)
Base	(48, 96, 192, 384, 512)	(3, 3, 3, 12, 3)	(3, 6, 12, 24, 32)
Large	(64, 128, 256, 512, 768)	(3, 3, 3, 12, 3)	(4, 8, 16, 32, 48)

Additionally, will smaller Utonia models be released in the future, such as small and tiny versions?

[Yujia-Zhang0913]

Hi.
The difference in model size is due to the fact that RoPE requires the number of channels to be divisible by 6.
As for the tiny/small versions, we may release the distill examples for reference in the future.

[Garfield2005]

Got it, thanks for the reply，looking forward to a smaller model.

Also, I’m trying to deploy the PTV model on Orin, but PTV3 cannot be converted to ONNX or TensorRT. @Gofinge mentioned in another issue that PTV4 is under development and will be easier to deploy. Do you have an estimated release timeline?

[Garfield2005]

Hi. The difference in model size is due to the fact that RoPE requires the number of channels to be divisible by 6. As for the tiny/small versions, we may release the distill examples for reference in the future.

Why not just use (48, 96, 192, 384, 512)? What considerations are there?

[Yujia-Zhang0913]

512 cannot be divided by 6.

[Garfield2005]

512 cannot be divided by 6.

Oh, sorry.

[niklas030]

Any Updates for the distilled version. I would love to start experimenting with the smaller models.