How do you design the embedding channels and head numbers

[Li-muye]

I'm trying to use another backbone which is not yolo family. And I find it difficult to modify the config, specifically, embedding channels and head numbers of neck and head, eg. embed_channels, in_channels, out_channels in neck. Are there any reasons to determine the parameters?

For example, I am using the config file yolo_world_v2_x_vlpan_bn_2e-3_100e_4x8gpus_obj365v1_goldg_train_lvis_minival.py.
The size of the feature embedding from the other backbone is [torch.Size(16, 96, 160, 160), torch.Size(16, 192, 80, 80), torch.Size(16, 384, 40, 40), torch.Size(16, 768, 20, 20)], while the feature from yolov8 (the original code) is [torch.Size([16, 320, 80, 80]), torch.Size([16, 640, 40, 40]), torch.Size([16, 640, 20, 20])], where 16 is the batch size in both cases and the image scale is [680,680].
embed_channels=[ 128, 256, 256, ], in_channels=[ 256, 512, 512, ], out_channels=[ 256, 512, 512, ], num_heads=[ 4, 8, 8, ]

how to modify the parameters or for what reasons these parameters are determined.