Other pre-trained E-RADIO variants?

[collinmccarthy]

Hello,

In radio/eradio_model.py I see quite a few model definitions for the hybrid/e-radio variants:

• fastervit2_large_fullres (ws7)
• fastervit2_large_fullres_ws8
• fastervit2_large_fullres_ws16
• fastervit2_large_fullres_ws32
• eradio_xxxtiny (ws16)
• eradio_xxxtiny_8x_ws12
• eradio_xxxtiny_8x_ws16

And then I see the eradio model is a wrapper around fastervit2_large_fullres_ws16, which matches what I see in radio/common.py.

Are there any other pre-trained e-radio models available right now or is it just this one "e-radio_v2"? I'm particularly interested in the (best performing) eradio_xxxtiny variant (probably one of the ws16 versions for a fair comparison). I'm assuming the older "eradio_v1" version is the same architecture as "e-radio_v2", but please correct me if I'm wrong.

Thanks!

[mranzinger]

Hi Collin,

So bad news/good news on this one. Unfortunately, the only pretrained E-RADIO we have right now is the e-radio_v2 one. However, we're putting together a training pass right now that will result in some new models. The first of which will be a ViT-B/16, most likely followed by a ViT-L/16, ViT-H/16, and a new E-RADIO. There was some demand from other people on an xxxtiny as well, so we can try to include it. What's your timeline for you work?

[collinmccarthy]

Hi Mike,

That sounds great, thank you. If a pre-trained xxxtiny model were available anytime in June, or even early-mid July, that would be a huge help for me personally. For now I can use the existing e-radio model. I'm assuming the new E-RADIO model will include an update to the source code / model definition so I can track down the changes that were made, but if not I would really appreciate that as well.

[Ali2500]

Hi, my query is kinda related to this so I'll just put it here: are there any particular issues with training E-RADIO using bfloat16? I want to fine-tune E-RADIO for downstream tasks using bfloat16 and I'm trying to figure out if this is feasible or not.

[gheinrich]

Hello, you should be good with E-RADIO and bfloat16. I did notice some issues with float16 (NaN) but none with bfloat16 and its wider range.