File size: 8,913 Bytes
ac6acf2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
import torch
import numpy as np
from typing import Union


def _to_tuple(x):
    if isinstance(x, int):
        return x, x
    else:
        return x


def get_fill_resize_and_crop(src, tgt):
    th, tw = _to_tuple(tgt)
    h, w = _to_tuple(src)

    tr = th / tw        # base resolution
    r = h / w           # target resolution

    # resize
    if r > tr:
        resize_height = th
        resize_width = int(round(th / h * w))
    else:
        resize_width = tw
        resize_height = int(round(tw / w * h))    # resize the target resolution down based on the base resolution

    crop_top = int(round((th - resize_height) / 2.0))
    crop_left = int(round((tw - resize_width) / 2.0))

    return (crop_top, crop_left), (crop_top + resize_height, crop_left + resize_width)


def get_meshgrid(start, *args):
    if len(args) == 0:
        # start is grid_size
        num = _to_tuple(start)
        start = (0, 0)
        stop = num
    elif len(args) == 1:
        # start is start, args[0] is stop, step is 1
        start = _to_tuple(start)
        stop = _to_tuple(args[0])
        num = (stop[0] - start[0], stop[1] - start[1])
    elif len(args) == 2:
        # start is start, args[0] is stop, args[1] is num
        start = _to_tuple(start)
        stop = _to_tuple(args[0])
        num = _to_tuple(args[1])
    else:
        raise ValueError(f"len(args) should be 0, 1 or 2, but got {len(args)}")

    grid_h = np.linspace(start[0], stop[0], num[0], endpoint=False, dtype=np.float32)
    grid_w = np.linspace(start[1], stop[1], num[1], endpoint=False, dtype=np.float32)
    grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    grid = np.stack(grid, axis=0)   # [2, W, H]
    return grid

#################################################################################
#                   Sine/Cosine Positional Embedding Functions                  #
#################################################################################
# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py

def get_2d_sincos_pos_embed(embed_dim, start, *args, cls_token=False, extra_tokens=0):
    """

    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 = get_meshgrid(start, *args)   # [2, H, w]
    # grid_h = np.arange(grid_size, dtype=np.float32)
    # grid_w = np.arange(grid_size, dtype=np.float32)
    # grid = np.meshgrid(grid_w, grid_h)  # here w goes first
    # grid = np.stack(grid, axis=0)   # [2, W, H]

    grid = grid.reshape([2, 1, *grid.shape[1:]])
    pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
    if cls_token and extra_tokens > 0:
        pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
    return pos_embed


def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
    assert embed_dim % 2 == 0

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

    emb = np.concatenate([emb_h, emb_w], axis=1)    # (H*W, 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 (W,H)

    out: (M, D)

    """
    assert embed_dim % 2 == 0
    omega = np.arange(embed_dim // 2, dtype=np.float64)
    omega /= embed_dim / 2.
    omega = 1. / 10000**omega  # (D/2,)

    pos = pos.reshape(-1)  # (M,)
    out = np.einsum('m,d->md', pos, omega)  # (M, D/2), outer product

    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


#################################################################################
#                   Rotary Positional Embedding Functions                       #
#################################################################################
# https://github.com/facebookresearch/llama/blob/main/llama/model.py#L443

def get_2d_rotary_pos_embed(embed_dim, start, *args, use_real=True):
    """

    This is a 2d version of precompute_freqs_cis, which is a RoPE for image tokens with 2d structure.



    Parameters

    ----------

    embed_dim: int

        embedding dimension size

    start: int or tuple of int

        If len(args) == 0, start is num; If len(args) == 1, start is start, args[0] is stop, step is 1;

        If len(args) == 2, start is start, args[0] is stop, args[1] is num.

    use_real: bool

        If True, return real part and imaginary part separately. Otherwise, return complex numbers.



    Returns

    -------

    pos_embed: torch.Tensor

        [HW, D/2]

    """
    grid = get_meshgrid(start, *args)   # [2, H, w]
    grid = grid.reshape([2, 1, *grid.shape[1:]])   # Returns a sampling matrix with the same resolution as the target resolution
    pos_embed = get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=use_real)
    return pos_embed


def get_2d_rotary_pos_embed_from_grid(embed_dim, grid, use_real=False):
    assert embed_dim % 4 == 0

    # use half of dimensions to encode grid_h
    emb_h = get_1d_rotary_pos_embed(embed_dim // 2, grid[0].reshape(-1), use_real=use_real)  # (H*W, D/4)
    emb_w = get_1d_rotary_pos_embed(embed_dim // 2, grid[1].reshape(-1), use_real=use_real)  # (H*W, D/4)

    if use_real:
        cos = torch.cat([emb_h[0], emb_w[0]], dim=1)    # (H*W, D/2)
        sin = torch.cat([emb_h[1], emb_w[1]], dim=1)    # (H*W, D/2)
        return cos, sin
    else:
        emb = torch.cat([emb_h, emb_w], dim=1)    # (H*W, D/2)
        return emb


def get_1d_rotary_pos_embed(dim: int, pos: Union[np.ndarray, int], theta: float = 10000.0, use_real=False):
    """

    Precompute the frequency tensor for complex exponentials (cis) with given dimensions.



    This function calculates a frequency tensor with complex exponentials using the given dimension 'dim'

    and the end index 'end'. The 'theta' parameter scales the frequencies.

    The returned tensor contains complex values in complex64 data type.



    Args:

        dim (int): Dimension of the frequency tensor.

        pos (np.ndarray, int): Position indices for the frequency tensor. [S] or scalar

        theta (float, optional): Scaling factor for frequency computation. Defaults to 10000.0.

        use_real (bool, optional): If True, return real part and imaginary part separately.

                                   Otherwise, return complex numbers.



    Returns:

        torch.Tensor: Precomputed frequency tensor with complex exponentials. [S, D/2]



    """
    if isinstance(pos, int):
        pos = np.arange(pos)
    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)[: (dim // 2)].float() / dim))  # [D/2]
    t = torch.from_numpy(pos).to(freqs.device)  # type: ignore  # [S]
    freqs = torch.outer(t, freqs).float()  # type: ignore   # [S, D/2]
    if use_real:
        freqs_cos = freqs.cos().repeat_interleave(2, dim=1)  # [S, D]
        freqs_sin = freqs.sin().repeat_interleave(2, dim=1)  # [S, D]
        return freqs_cos, freqs_sin
    else:
        freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64     # [S, D/2]
        return freqs_cis



def calc_sizes(rope_img, patch_size, th, tw):
    if rope_img == 'extend':
        # Expansion mode
        sub_args = [(th, tw)]
    elif rope_img.startswith('base'):
        # Based on the specified dimensions, other dimensions are obtained through interpolation.
        base_size = int(rope_img[4:]) // 8 // patch_size
        start, stop = get_fill_resize_and_crop((th, tw), base_size)
        sub_args = [start, stop, (th, tw)]
    else:
        raise ValueError(f"Unknown rope_img: {rope_img}")
    return sub_args


def init_image_posemb(rope_img,

                      resolutions,

                      patch_size,

                      hidden_size,

                      num_heads,

                      log_fn,

                      rope_real=True,

                      ):
    freqs_cis_img = {}
    for reso in resolutions:
        th, tw = reso.height // 8 // patch_size, reso.width // 8 // patch_size
        sub_args = calc_sizes(rope_img, patch_size, th, tw)
        freqs_cis_img[str(reso)] = get_2d_rotary_pos_embed(hidden_size // num_heads, *sub_args, use_real=rope_real)
        log_fn(f"    Using image RoPE ({rope_img}) ({'real' if rope_real else 'complex'}): {sub_args} | ({reso}) "
               f"{freqs_cis_img[str(reso)][0].shape if rope_real else freqs_cis_img[str(reso)].shape}")
    return freqs_cis_img