Spaces:
Running on Zero
Running on Zero
File size: 7,639 Bytes
04012a7 | 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 | import math
from typing import Any, Optional, Tuple
import numpy as np
import torch
from torch import nn
class PositionEmbeddingSine(nn.Module):
def __init__(
self,
num_pos_feats,
temperature: int = 10000,
normalize: bool = True,
scale: Optional[float] = None,
warmup_cache: bool = True,
image_size: int = 1024,
strides: Tuple[int] = (4, 8, 16, 32),
):
super().__init__()
assert num_pos_feats % 2 == 0, "Expecting even model width"
self.num_pos_feats = num_pos_feats // 2
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
self.cache = {}
if warmup_cache:
device = torch.device("cpu")
for stride in strides:
cache_key = (image_size // stride, image_size // stride)
self._pe(1, device, *cache_key)
def _encode_xy(self, x, y):
assert len(x) == len(y) and x.ndim == y.ndim == 1
x_embed = x * self.scale
y_embed = y * 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=2
).flatten(1)
pos_y = torch.stack(
(pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
).flatten(1)
return pos_x, pos_y
@torch.no_grad()
def encode_boxes(self, x, y, w, h):
pos_x, pos_y = self._encode_xy(x, y)
pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
return pos
encode = encode_boxes
@torch.no_grad()
def encode_points(self, x, y, labels):
(bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
assert bx == by and nx == ny and bx == bl and nx == nl
pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
return pos
@torch.no_grad()
def _pe(self, B, device, *cache_key):
H, W = cache_key
if cache_key in self.cache:
return self.cache[cache_key].to(device)[None].repeat(B, 1, 1, 1)
y_embed = (
torch.arange(1, H + 1, dtype=torch.float32, device=device)
.view(1, -1, 1)
.repeat(B, 1, W)
)
x_embed = (
torch.arange(1, W + 1, dtype=torch.float32, device=device)
.view(1, 1, -1)
.repeat(B, H, 1)
)
if self.normalize:
eps = 1e-6
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=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)
self.cache[cache_key] = pos[0]
return pos
@torch.no_grad()
def forward(self, x: torch.Tensor):
B = x.shape[0]
cache_key = (x.shape[-2], x.shape[-1])
return self._pe(B, x.device, *cache_key)
class PositionEmbeddingRandom(nn.Module):
def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
super().__init__()
if scale is None or scale <= 0.0:
scale = 1.0
self.register_buffer(
"positional_encoding_gaussian_matrix",
scale * torch.randn((2, num_pos_feats)),
)
def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
coords = 2 * coords - 1
coords = coords @ self.positional_encoding_gaussian_matrix
coords = 2 * np.pi * coords
return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
def forward(self, size: Tuple[int, int]) -> torch.Tensor:
h, w = size
device: Any = self.positional_encoding_gaussian_matrix.device
grid = torch.ones((h, w), device=device, dtype=torch.float32)
y_embed = grid.cumsum(dim=0) - 0.5
x_embed = grid.cumsum(dim=1) - 0.5
y_embed = y_embed / h
x_embed = x_embed / w
pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
return pe.permute(2, 0, 1)
def forward_with_coords(
self, coords_input: torch.Tensor, image_size: Tuple[int, int]
) -> torch.Tensor:
coords = coords_input.clone()
coords[:, :, 0] = coords[:, :, 0] / image_size[1]
coords[:, :, 1] = coords[:, :, 1] / image_size[0]
return self._pe_encoding(coords.to(torch.float))
def init_t_xy(end_x: int, end_y: int):
t = torch.arange(end_x * end_y, dtype=torch.float32)
t_x = (t % end_x).float()
t_y = torch.div(t, end_x, rounding_mode="floor").float()
return t_x, t_y
def compute_axial_cis(dim: int, end_x: int, end_y: int, theta: float = 10000.0):
freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
t_x, t_y = init_t_xy(end_x, end_y)
freqs_x = torch.outer(t_x, freqs_x)
freqs_y = torch.outer(t_y, freqs_y)
freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)
def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
ndim = x.ndim
assert 0 <= 1 < ndim
assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
return freqs_cis.view(*shape)
def apply_rotary_enc(
xq: torch.Tensor,
xk: torch.Tensor,
freqs_cis: torch.Tensor,
repeat_freqs_k: bool = False,
):
xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
xk_ = (
torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
if xk.shape[-2] != 0
else None
)
freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
if xk_ is None:
return xq_out.type_as(xq).to(xq.device), xk
if repeat_freqs_k:
r = xk_.shape[-2] // xq_.shape[-2]
if freqs_cis.is_cuda:
freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
else:
freqs_cis = freqs_cis.unsqueeze(2).expand(-1, -1, r, -1, -1).flatten(2, 3)
xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)
|