Varrying Feature sizes for Anchors

[jaideep11061982]

Is your feature request related to a problem? Please describe.
A clear and concise description of what the problem is.
In Anchors.py
Currently for all the levels we build the feature size as below

``def get_feat_sizes(image_size: Tuple[int, int], max_level: int): """Get feat widths and heights for all levels. Args: image_size: a tuple (H, W) max_level: maximum feature level. Returns: feat_sizes: a list of tuples (height, width) for each level. """ feat_size = image_size feat_sizes = [feat_size] for _ in range(1, max_level + 1): feat_size = ((feat_size[0] - 1) // 2 + 1, (feat_size[1] - 1) // 2 + 1) feat_sizes.append(feat_size) return feat_sizes
Is itpossile to build features of varying size like //2,//3//4 etc . like in Yolo

Describe the solution you'd like
A clear and concise description of what you want to happen.

Describe alternatives you've considered
A clear and concise description of any alternative solutions or features you've considered.

Additional context
Add any other context or screenshots about the feature request here.

[rwightman]

@jaideep11061982 converted to discussion so it lives longer..

As per #220, you can try adjusting via the anchor scale config, if specified as a tuple it will be applied per feature level ... this code based on TF models retinanet anchors so you can also dig through there for further examples .. although warning, it's a lot to dig through.

[jaideep11061982]

Thanks Ross. Just for every one awareness. Yes studied and I tried that , but pretrained wts load fails , I will have to apply weight filter function to succeed it through .one more restriction is I cant increase length of aspect ratio or else again weight load fails . Some more question , 1. why we have tuple of aspect ratio which are inverse of each instead of single element sequence ,i tried to understand but couldn't well . 2. Based on aspect ratio stats of data I see that there are quite a number of bounding boxes that have aspect ratio above 2. So default aspect ratio tuple be able to catch that ?

…

On Sun, 13 Jun 2021, 06:13 Ross Wightman, ***@***.***> wrote: @jaideep11061982 <https://github.com/jaideep11061982> converted to discussion so it lives longer.. As per #220 <#220>, you can try adjusting via the anchor scale config, if specified as a tuple it will be applied per feature level ... this code based on TF models retinanet anchors so you can also dig through there for further examples .. although warning, it's a lot to dig through. — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#225 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AJDFNZGH6PYCB2ZJWNELBZ3TSP5KDANCNFSM46THFDAQ> .