minerva.utils.position_embedding

Functions

get_1d_sincos_pos_embed_from_grid(embed_dim, pos)

embed_dim: output dimension for each position

get_2d_sincos_pos_embed(embed_dim, grid_size[, cls_token])

grid_size: int or tuple/list of (grid_h, grid_w)

get_2d_sincos_pos_embed_from_grid(embed_dim, grid)

interpolate_pos_embed(model, checkpoint_model[, ...])

Module Contents

minerva.utils.position_embedding.get_1d_sincos_pos_embed_from_grid(embed_dim, pos)[source]

embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)

minerva.utils.position_embedding.get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False)[source]
grid_size: int or tuple/list of (grid_h, grid_w)

  • If int, the grid is square

  • If tuple/list, grid_h and grid_w define a rectangular grid

return:

pos_embed: [grid_h*grid_w, embed_dim] or [1+grid_h*grid_w, embed_dim] (w/ cls_token)

minerva.utils.position_embedding.get_2d_sincos_pos_embed_from_grid(embed_dim, grid)[source]
minerva.utils.position_embedding.interpolate_pos_embed(model, checkpoint_model, newsize1=None, newsize2=None)[source]