models/position_encoding.py

"""
Various positional encodings for the transformer.
Modified from DETR (https://github.com/facebookresearch/detr)
"""
import math
import torch
from torch import nn

from util.misc import NestedTensor

# dimension == 1
class PositionEmbeddingSine1D(nn.Module):
    """
    This is a more standard version of the position embedding, very similar to the one
    used by the Attention is all you need paper, generalized to work on images.
    """
    def __init__(self, num_pos_feats=256, temperature=10000, normalize=False, scale=None):
        super().__init__()
        self.num_pos_feats = num_pos_feats
        self.temperature = temperature
        self.normalize = normalize
        if scale is not None and normalize is False:
            raise ValueError("normalize should be True if scale is passed")
        if scale is None:
            scale = 2 * math.pi
        self.scale = scale

    def forward(self, tensor_list: NestedTensor):
        x = tensor_list.tensors # [B, C, T]
        mask = tensor_list.mask # [B, T]
        assert mask is not None
        not_mask = ~mask
        x_embed = not_mask.cumsum(1, dtype=torch.float32)  # [B, T]
        if self.normalize:
            eps = 1e-6
            x_embed = x_embed / (x_embed[:, -1:] + eps) * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)

        pos_x = x_embed[:, :, None] / dim_t  # [B, T, C]
        # n,c,t
        pos_x = torch.stack((pos_x[:, :, 0::2].sin(), pos_x[:, :, 1::2].cos()), dim=3).flatten(2)
        pos = pos_x.permute(0, 2, 1)    # [B, C, T]
        return pos

# dimension == 2
class PositionEmbeddingSine2D(nn.Module):
    """
    This is a more standard version of the position embedding, very similar to the one
    used by the Attention is all you need paper, generalized to work on images.
    """
    def __init__(self, num_pos_feats=64, temperature=10000, normalize=False, scale=None):
        super().__init__()
        self.num_pos_feats = num_pos_feats
        self.temperature = temperature
        self.normalize = normalize
        if scale is not None and normalize is False:
            raise ValueError("normalize should be True if scale is passed")
        if scale is None:
            scale = 2 * math.pi
        self.scale = scale

    def forward(self, tensor_list: NestedTensor):
        x = tensor_list.tensors  # [B, C, H, W]
        mask = tensor_list.mask  # [B, H, W]
        assert mask is not None
        not_mask = ~mask
        y_embed = not_mask.cumsum(1, dtype=torch.float32)
        x_embed = not_mask.cumsum(2, dtype=torch.float32)
        if self.normalize:
            eps = 1e-6
            y_embed = (y_embed - 0.5) / (y_embed[:, -1:, :] + eps) * self.scale
            x_embed = (x_embed - 0.5) / (x_embed[:, :, -1:] + eps) * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_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)
        return pos  # [B, C, H, W]


# dimension == 3
class PositionEmbeddingSine3D(nn.Module):
    """
    This is a more standard version of the position embedding, very similar to the one
    used by the Attention is all you need paper, generalized to work on images.
    """
    def __init__(self, num_pos_feats=64, num_frames=36, temperature=10000, normalize=False, scale=None):
        super().__init__()
        self.num_pos_feats = num_pos_feats
        self.temperature = temperature
        self.normalize = normalize
        self.frames = num_frames
        if scale is not None and normalize is False:
            raise ValueError("normalize should be True if scale is passed")
        if scale is None:
            scale = 2 * math.pi
        self.scale = scale

    def forward(self, tensor_list: NestedTensor):
        x = tensor_list.tensors # [B*T, C, H, W]
        mask = tensor_list.mask # [B*T, H, W]
        n,h,w = mask.shape
        mask = mask.reshape(n//self.frames, self.frames,h,w) # [B, T, H, W]
        assert mask is not None
        not_mask = ~mask
        z_embed = not_mask.cumsum(1, dtype=torch.float32) # [B, T, H, W]
        y_embed = not_mask.cumsum(2, dtype=torch.float32) # [B, T, H, W]
        x_embed = not_mask.cumsum(3, dtype=torch.float32) # [B, T, H, W]
        if self.normalize:
            eps = 1e-6
            z_embed = z_embed / (z_embed[:, -1:, :, :] + eps) * self.scale
            y_embed = y_embed / (y_embed[:, :, -1:, :] + eps) * self.scale
            x_embed = x_embed / (x_embed[:, :, :, -1:] + eps) * self.scale

        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device) # 
        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)

        pos_x = x_embed[:, :, :, :, None] / dim_t # [B, T, H, W, c]
        pos_y = y_embed[:, :, :, :, None] / dim_t
        pos_z = z_embed[:, :, :, :, None] / dim_t
        pos_x = torch.stack((pos_x[:, :, :, :, 0::2].sin(), pos_x[:, :, :, :, 1::2].cos()), dim=5).flatten(4) # [B, T, H, W, c]
        pos_y = torch.stack((pos_y[:, :, :, :, 0::2].sin(), pos_y[:, :, :, :, 1::2].cos()), dim=5).flatten(4)
        pos_z = torch.stack((pos_z[:, :, :, :, 0::2].sin(), pos_z[:, :, :, :, 1::2].cos()), dim=5).flatten(4)
        pos = torch.cat((pos_z, pos_y, pos_x), dim=4).permute(0, 1, 4, 2, 3) # [B, T, C, H, W]
        return pos



def build_position_encoding(args):
    # build 2D position encoding
    N_steps = args.hidden_dim // 2  # 256 / 2 = 128
    if args.position_embedding in ('v2', 'sine'):
        # TODO find a better way of exposing other arguments
        position_embedding = PositionEmbeddingSine2D(N_steps, normalize=True)
    else:
        raise ValueError(f"not supported {args.position_embedding}")

    return position_embedding