Skip to content

deeplabcut.pose_estimation_pytorch.models.necks.utils

Functions:

Name Description
make_sine_position_embedding

Generate sine position embeddings for a given height, width, and model dimension.

make_sine_position_embedding 

make_sine_position_embedding(
    h: int, w: int, d_model: int, temperature: int = 10000, scale: float = 2 * math.pi
) -> torch.Tensor

Generate sine position embeddings for a given height, width, and model dimension.

Parameters:

Name Type Description Default

h

int

Height of the embedding.

 required

w

int

Width of the embedding.

 required

d_model

int

Dimension of the model.

 required

temperature

int

Temperature parameter for position embedding calculation. Defaults to 10000.

 10000

scale

float

Scaling factor for position embedding. Defaults to 2 * math.pi.

 2 * pi

Returns:

Type Description
Tensor

Sine position embeddings with shape (batch_size, d_model, h * w).
Example

h, w, d_model = 10, 20, 512 pos_emb = make_sine_position_embedding(h, w, d_model) print(pos_emb.shape) # Output: torch.Size([1, 512, 200])

Source code in deeplabcut/pose_estimation_pytorch/models/necks/utils.py 
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
def make_sine_position_embedding(
    h: int, w: int, d_model: int, temperature: int = 10000, scale: float = 2 * math.pi
) -> torch.Tensor:
    """Generate sine position embeddings for a given height, width, and model dimension.

    Args:
        h: Height of the embedding.
        w: Width of the embedding.
        d_model: Dimension of the model.
        temperature: Temperature parameter for position embedding calculation.
                     Defaults to 10000.
        scale: Scaling factor for position embedding. Defaults to 2 * math.pi.

    Returns:
        Sine position embeddings with shape (batch_size, d_model, h * w).

    Example:
        >>> h, w, d_model = 10, 20, 512
        >>> pos_emb = make_sine_position_embedding(h, w, d_model)
        >>> print(pos_emb.shape)  # Output: torch.Size([1, 512, 200])
    """
    area = torch.ones(1, h, w)
    y_embed = area.cumsum(1, dtype=torch.float32)
    x_embed = area.cumsum(2, dtype=torch.float32)
    one_direction_feats = d_model // 2
    eps = 1e-6
    y_embed = y_embed / (y_embed[:, -1:, :] + eps) * scale
    x_embed = x_embed / (x_embed[:, :, -1:] + eps) * scale

    dim_t = torch.arange(one_direction_feats, dtype=torch.float32)
    dim_t = temperature ** (2 * (dim_t // 2) / one_direction_feats)

    pos_x = x_embed[:, :, :, None] / dim_t
    pos_y = y_embed[:, :, :, None] / dim_t
    pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3)
    pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3)
    pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
    pos = pos.flatten(2).permute(0, 2, 1)

    return pos