File size: 11,108 Bytes
fcefac1 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 | # Copyright (c) Meta Platforms, Inc. and affiliates.
#
# This software may be used and distributed in accordance with
# the terms of the DINOv3 License Agreement.
import math
from typing import Literal
import numpy as np
import torch
from torch import Tensor, nn
# RoPE positional embedding with no mixing of coordinates (axial) and no learnable weights
# Supports two parametrizations of the rope parameters: either using `base` or `min_period` and `max_period`.
class RopePositionEmbedding(nn.Module):
def __init__(
self,
embed_dim: int,
*,
num_heads: int,
base: float | None = 100.0,
min_period: float | None = None,
max_period: float | None = None,
normalize_coords: Literal["min", "max", "separate"] = "separate",
shift_coords: float | None = None,
jitter_coords: float | None = None,
rescale_coords: float | None = None,
dtype: torch.dtype | None = None,
device: torch.device | None = None,
):
super().__init__()
assert embed_dim % (4 * num_heads) == 0
both_periods = min_period is not None and max_period is not None
if (base is None and not both_periods) or (base is not None and both_periods):
raise ValueError("Either `base` or `min_period`+`max_period` must be provided.")
D_head = embed_dim // num_heads
self.base = base
self.min_period = min_period
self.max_period = max_period
self.D_head = D_head
self.normalize_coords = normalize_coords
self.shift_coords = shift_coords
self.jitter_coords = jitter_coords
self.rescale_coords = rescale_coords
# Needs persistent=True because we do teacher.load_state_dict(student.state_dict()) to initialize the teacher
self.dtype = dtype # Don't rely on self.periods.dtype
self.register_buffer(
"periods",
torch.empty(D_head // 4, device=device, dtype=dtype),
persistent=True,
)
self._init_weights()
def forward(self, *, H: int, W: int) -> tuple[Tensor, Tensor]:
device = self.periods.device
dtype = self.dtype
dd = {"device": device, "dtype": dtype}
# Prepare coords in range [-1, +1]
if self.normalize_coords == "max":
max_HW = max(H, W)
coords_h = torch.arange(0.5, H, **dd) / max_HW # [H]
coords_w = torch.arange(0.5, W, **dd) / max_HW # [W]
elif self.normalize_coords == "min":
min_HW = min(H, W)
coords_h = torch.arange(0.5, H, **dd) / min_HW # [H]
coords_w = torch.arange(0.5, W, **dd) / min_HW # [W]
elif self.normalize_coords == "separate":
coords_h = torch.arange(0.5, H, **dd) / H # [H]
coords_w = torch.arange(0.5, W, **dd) / W # [W]
else:
raise ValueError(f"Unknown normalize_coords: {self.normalize_coords}")
coords = torch.stack(torch.meshgrid(coords_h, coords_w, indexing="ij"), dim=-1) # [H, W, 2]
coords = coords.flatten(0, 1) # [HW, 2]
coords = 2.0 * coords - 1.0 # Shift range [0, 1] to [-1, +1]
# Shift coords by adding a uniform value in [-shift, shift]
if self.training and self.shift_coords is not None:
shift_hw = torch.empty(2, **dd).uniform_(-self.shift_coords, self.shift_coords)
coords += shift_hw[None, :]
# Jitter coords by multiplying the range [-1, 1] by a log-uniform value in [1/jitter, jitter]
if self.training and self.jitter_coords is not None:
jitter_max = np.log(self.jitter_coords)
jitter_min = -jitter_max
jitter_hw = torch.empty(2, **dd).uniform_(jitter_min, jitter_max).exp()
coords *= jitter_hw[None, :]
# Rescale coords by multiplying the range [-1, 1] by a log-uniform value in [1/rescale, rescale]
if self.training and self.rescale_coords is not None:
rescale_max = np.log(self.rescale_coords)
rescale_min = -rescale_max
rescale_hw = torch.empty(1, **dd).uniform_(rescale_min, rescale_max).exp()
coords *= rescale_hw
# Prepare angles and sin/cos
angles = 2 * math.pi * coords[:, :, None] / self.periods[None, None, :] # [HW, 2, D//4]
angles = angles.flatten(1, 2) # [HW, D//2]
angles = angles.tile(2) # [HW, D]
cos = torch.cos(angles) # [HW, D]
sin = torch.sin(angles) # [HW, D]
return (sin, cos) # 2 * [HW, D]
def _init_weights(self):
device = self.periods.device
dtype = self.dtype
if self.base is not None:
periods = self.base ** (
2 * torch.arange(self.D_head // 4, device=device, dtype=dtype) / (self.D_head // 2)
) # [D//4]
else:
base = self.max_period / self.min_period
exponents = torch.linspace(0, 1, self.D_head // 4, device=device, dtype=dtype) # [D//4] range [0, 1]
periods = base**exponents # range [1, max_period / min_period]
periods = periods / base # range [min_period / max_period, 1]
periods = periods * self.max_period # range [min_period, max_period]
self.periods.data = periods
class RopePositionEmbedding3D(nn.Module):
"""
RoPE positional embedding for 3D grids with no mixing across axes (axial),
and no learnable weights.
Supports two parametrizations of the rope parameters: either using `base`
or `min_period` + `max_period`.
Returns (sin, cos) each shaped [D*H*W, D_head], suitable for applying to q/k.
"""
def __init__(
self,
embed_dim: int,
*,
num_heads: int,
base: float | None = 100.0,
min_period: float | None = None,
max_period: float | None = None,
normalize_coords: Literal["min", "max", "separate"] = "separate",
shift_coords: float | None = None,
jitter_coords: float | None = None,
rescale_coords: float | None = None,
dtype: torch.dtype | None = None,
device: torch.device | None = None,
):
super().__init__()
# For 3 axes, per-axis rotary block is D_head//6
assert embed_dim % (6 * num_heads) == 0, \
"For 3D RoPE, embed_dim must be divisible by 6 * num_heads."
both_periods = (min_period is not None) and (max_period is not None)
if (base is None and not both_periods) or (base is not None and both_periods):
raise ValueError("Either `base` or `min_period`+`max_period` must be provided.")
D_head = embed_dim // num_heads
self.base = base
self.min_period = min_period
self.max_period = max_period
self.D_head = D_head
self.normalize_coords = normalize_coords
self.shift_coords = shift_coords
self.jitter_coords = jitter_coords
self.rescale_coords = rescale_coords
# Keep a persistent buffer so teacher.load_state_dict(student.state_dict()) works.
self.dtype = dtype # Don't rely on self.periods.dtype
self.register_buffer(
"periods",
torch.empty(D_head // 6, device=device, dtype=dtype),
persistent=True,
)
self._init_weights()
def _init_weights(self):
device = self.periods.device
dtype = self.dtype
per_axis_block = self.D_head // 6 # 3 axes × per_axis_block = D_head//2
if self.base is not None:
# Classic exponential spectrum
periods = self.base ** (
2 * torch.arange(per_axis_block, device=device, dtype=dtype) / (self.D_head // 2)
) # [per_axis_block]
else:
# Linearly spaced in log-period between min_period and max_period
base = self.max_period / self.min_period
exponents = torch.linspace(0, 1, per_axis_block, device=device, dtype=dtype)
periods = base**exponents # [1, base]
periods = periods / base # [1/base, 1]
periods = periods * self.max_period # [min_period, max_period]
self.periods.data = periods
def forward(self, *, D: int, H: int, W: int) -> tuple[Tensor, Tensor]:
"""
Args:
D, H, W: depth, height, width (integers)
Returns:
(sin, cos): two tensors of shape [D*H*W, D_head]
"""
device = self.periods.device
dtype = self.dtype
dd = {"device": device, "dtype": dtype}
# --- Build normalized coords in [-1, +1] for each axis ---
if self.normalize_coords == "max":
m = max(D, H, W)
coords_d = torch.arange(0.5, D, **dd) / m # [D]
coords_h = torch.arange(0.5, H, **dd) / m # [H]
coords_w = torch.arange(0.5, W, **dd) / m # [W]
elif self.normalize_coords == "min":
m = min(D, H, W)
coords_d = torch.arange(0.5, D, **dd) / m
coords_h = torch.arange(0.5, H, **dd) / m
coords_w = torch.arange(0.5, W, **dd) / m
elif self.normalize_coords == "separate":
coords_d = torch.arange(0.5, D, **dd) / D
coords_h = torch.arange(0.5, H, **dd) / H
coords_w = torch.arange(0.5, W, **dd) / W
else:
raise ValueError(f"Unknown normalize_coords: {self.normalize_coords}")
# Meshgrid in (D, H, W) order; last dim stacks (d, h, w)
coords = torch.stack(
torch.meshgrid(coords_d, coords_h, coords_w, indexing="ij"),
dim=-1
) # [D, H, W, 3]
coords = coords.flatten(0, 2) # [DHW, 3]
coords = 2.0 * coords - 1.0 # shift [0,1] -> [-1,1]
# --- Optional data-aug on coords (train-time only) ---
if self.training and self.shift_coords is not None:
shift_dhw = torch.empty(3, **dd).uniform_(-self.shift_coords, self.shift_coords)
coords = coords + shift_dhw[None, :]
if self.training and self.jitter_coords is not None:
jitter_max = np.log(self.jitter_coords)
jitter_min = -jitter_max
jitter_dhw = torch.empty(3, **dd).uniform_(jitter_min, jitter_max).exp()
coords = coords * jitter_dhw[None, :]
if self.training and self.rescale_coords is not None:
rescale_max = np.log(self.rescale_coords)
rescale_min = -rescale_max
rescale = torch.empty(1, **dd).uniform_(rescale_min, rescale_max).exp()
coords = coords * rescale
# --- Build rotary angles (axial, no mixing) ---
# periods: [P] where P = D_head//6
# coords: [N, 3], broadcast => [N, 3, P]
angles = 2 * math.pi * coords[:, :, None] / self.periods[None, None, :] # [N, 3, P]
angles = angles.flatten(1, 2) # [N, 3*P] == [N, D_head//2]
angles = angles.tile(2) # [N, D_head]
cos = torch.cos(angles)
sin = torch.sin(angles)
return (sin, cos) |