Question about relative_coords_table buffer in Swin V2.

[BlueDruddigon]

In Swin Transformer V2, as proposed using Log-Space CPB,

In the code implementation, log-spaced coordinates are generated when transferring across window sizes through relative_coords_table like in this code snippet.

Swin-Transformer/models/swin_transformer_v2.py

Lines 109 to 111 in f92123a

	relative_coords_table *= 8 # normalize to -8, 8
	relative_coords_table = torch.sign(relative_coords_table) * torch.log2(
	torch.abs(relative_coords_table) + 1.0) / np.log2(8)

I have two questions:

• Does it mean to be 8 and np.log2(8) ? Why do you do that ?
• And if I do not multiply relative_coords_table by 8 and remove the afterwards divided by np.log2(8), what has changed?