File size: 4,471 Bytes
86b1a7e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
325137b
 
 
86b1a7e
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
325137b
 
 
86b1a7e
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
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
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
# Adapted from: https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/embeddings.py
import math

import numpy as np
import torch
from einops import rearrange
from torch import nn


def get_timestep_embedding(
    timesteps: torch.Tensor,
    embedding_dim: int,
    flip_sin_to_cos: bool = False,
    downscale_freq_shift: float = 1,
    scale: float = 1,
    max_period: int = 10000,
):
    """
    This matches the implementation in Denoising Diffusion Probabilistic Models: Create sinusoidal timestep embeddings.

    :param timesteps: a 1-D Tensor of N indices, one per batch element.
                      These may be fractional.
    :param embedding_dim: the dimension of the output. :param max_period: controls the minimum frequency of the
    embeddings. :return: an [N x dim] Tensor of positional embeddings.
    """
    assert len(timesteps.shape) == 1, "Timesteps should be a 1d-array"

    half_dim = embedding_dim // 2
    exponent = -math.log(max_period) * torch.arange(
        start=0, end=half_dim, dtype=torch.float32, device=timesteps.device
    )
    exponent = exponent / (half_dim - downscale_freq_shift)

    emb = torch.exp(exponent)
    emb = timesteps[:, None].float() * emb[None, :]

    # scale embeddings
    emb = scale * emb

    # concat sine and cosine embeddings
    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)

    # flip sine and cosine embeddings
    if flip_sin_to_cos:
        emb = torch.cat([emb[:, half_dim:], emb[:, :half_dim]], dim=-1)

    # zero pad
    if embedding_dim % 2 == 1:
        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
    return emb


def get_3d_sincos_pos_embed(embed_dim, grid, w, h, f):
    """
    grid_size: int of the grid height and width return: pos_embed: [grid_size*grid_size, embed_dim] or
    [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
    """
    grid = rearrange(grid, "c (f h w) -> c f h w", h=h, w=w)
    grid = rearrange(grid, "c f h w -> c h w f", h=h, w=w)
    grid = grid.reshape([3, 1, w, h, f])
    pos_embed = get_3d_sincos_pos_embed_from_grid(embed_dim, grid)
    pos_embed = pos_embed.transpose(1, 0, 2, 3)
    return rearrange(pos_embed, "h w f c -> (f h w) c")


def get_3d_sincos_pos_embed_from_grid(embed_dim, grid):
    if embed_dim % 3 != 0:
        raise ValueError("embed_dim must be divisible by 3")

    # use half of dimensions to encode grid_h
    emb_f = get_1d_sincos_pos_embed_from_grid(embed_dim // 3, grid[0])  # (H*W*T, D/3)
    emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 3, grid[1])  # (H*W*T, D/3)
    emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 3, grid[2])  # (H*W*T, D/3)

    emb = np.concatenate([emb_h, emb_w, emb_f], axis=-1)  # (H*W*T, D)
    return emb


def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
    """
    embed_dim: output dimension for each position pos: a list of positions to be encoded: size (M,) out: (M, D)
    """
    if embed_dim % 2 != 0:
        raise ValueError("embed_dim must be divisible by 2")

    omega = np.arange(embed_dim // 2, dtype=np.float64)
    omega /= embed_dim / 2.0
    omega = 1.0 / 10000**omega  # (D/2,)

    pos_shape = pos.shape

    pos = pos.reshape(-1)
    out = np.einsum("m,d->md", pos, omega)  # (M, D/2), outer product
    out = out.reshape([*pos_shape, -1])[0]

    emb_sin = np.sin(out)  # (M, D/2)
    emb_cos = np.cos(out)  # (M, D/2)

    emb = np.concatenate([emb_sin, emb_cos], axis=-1)  # (M, D)
    return emb


class SinusoidalPositionalEmbedding(nn.Module):
    """Apply positional information to a sequence of embeddings.

    Takes in a sequence of embeddings with shape (batch_size, seq_length, embed_dim) and adds positional embeddings to
    them

    Args:
        embed_dim: (int): Dimension of the positional embedding.
        max_seq_length: Maximum sequence length to apply positional embeddings

    """

    def __init__(self, embed_dim: int, max_seq_length: int = 32):
        super().__init__()
        position = torch.arange(max_seq_length).unsqueeze(1)
        div_term = torch.exp(
            torch.arange(0, embed_dim, 2) * (-math.log(10000.0) / embed_dim)
        )
        pe = torch.zeros(1, max_seq_length, embed_dim)
        pe[0, :, 0::2] = torch.sin(position * div_term)
        pe[0, :, 1::2] = torch.cos(position * div_term)
        self.register_buffer("pe", pe)

    def forward(self, x):
        _, seq_length, _ = x.shape
        x = x + self.pe[:, :seq_length]
        return x