File size: 12,636 Bytes
d80b3db | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# Copyright 2025 AI for Oncology Research Group. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
#
# References:
# https://github.com/facebookresearch/dino/blob/main/vision_transformer.py
# https://github.com/rwightman/pytorch-image-models/tree/master/timm/models/vision_transformer.py
from functools import partial
from typing import Callable
from typing_extensions import override
import torch
from torch import nn
from .attention import MemEffAttention
from .transformer_block import NestedTensorBlock as Block
from .vision_transformer_base import (
DinoVisionTransformerBase,
DinoVisionTransformerDim,
DinoVisionTransformerFFNLayer,
)
class DinoVisionTransformer(DinoVisionTransformerBase):
"""DinoVisionTransformer for 2D images.
Parameters
----------
img_size : int or tuple[int, int]
Input image size, either a single integer or a tuple of two integers (height, width).
patch_size : int or tuple[int, int]
Patch size, either a single integer or a tuple of two integers (height, width).
in_chans : int
Number of input channels, default is 3.
embed_dim : int
Embedding dimension.
depth : int
Depth of transformer.
num_heads : int
Number of attention heads.
mlp_ratio : int
Ratio of mlp hidden dim to embedding dim.
qkv_bias : bool
Enable bias for qkv if True.
proj_bias : bool
Enable bias for proj in attn if True.
ffn_bias : bool
Enable bias for ffn if True.
drop_path_rate : float
Stochastic depth rate.
drop_path_uniform : bool
Apply uniform drop rate across blocks.
weight_init : str
Weight init scheme.
init_values : float
Layer-scale init values.
act_layer : nn.Module
MLP activation layer.
block_fn : nn.Module
Transformer block class.
ffn_layer : DinoVisionTransformerFFNLayer
Type of FFN layer to use, can be DinoVisionTransformerFFNLayer.MLP,
DinoVisionTransformerFFNLayer.SWIGLU, DinoVisionTransformerFFNLayer.SWIGLU_FUSED,
or DinoVisionTransformerFFNLayer.IDENTITY. Default is DinoVisionTransformerFFNLayer.MLP.
block_chunks : int
Split block sequence into block_chunks units for FSDP wrap.
num_register_tokens : int
Number of extra tokens for the model to deposit information (so-called "registers"). Default is 0.
interpolate_antialias : str
Flag to apply anti-aliasing when interpolating positional embeddings.
interpolate_offset : float
Work-around offset to apply when interpolating positional embeddings.
"""
def __init__(
self,
img_size: int | tuple[int, int] = 224,
patch_size: int | tuple[int, int] = 16,
in_chans: int = 3,
embed_dim: int = 768,
depth: int = 12,
num_heads: int = 12,
mlp_ratio: float = 4.0,
qkv_bias: bool = True,
ffn_bias: bool = True,
proj_bias: bool = True,
drop_path_rate: float = 0.0,
drop_path_uniform: bool = False,
init_values: float | None = None, # for layerscale: None or 0 => no layerscale
act_layer: Callable[..., nn.Module] = nn.GELU,
block_fn: Callable[..., Block] = Block,
ffn_layer: DinoVisionTransformerFFNLayer = DinoVisionTransformerFFNLayer.MLP,
block_chunks: int = 1,
num_register_tokens: int = 0,
interpolate_antialias: bool = False,
interpolate_offset: float = 0.1,
) -> None:
"""Inits :class:`DinoVisionTransformer`.
Parameters
----------
img_size : int or tuple[int, int]
Input image size, either a single integer or a tuple of two integers (height, width).
patch_size : int or tuple[int, int]
Patch size, either a single integer or a tuple of two integers (height, width).
in_chans : int
Number of input channels, default is 3.
embed_dim : int
Embedding dimension.
depth : int
Depth of transformer.
num_heads : int
Number of attention heads.
mlp_ratio : int
Ratio of mlp hidden dim to embedding dim.
qkv_bias : bool
Enable bias for qkv if True.
proj_bias : bool
Enable bias for proj in attn if True.
ffn_bias : bool
Enable bias for ffn if True.
drop_path_rate : float
Stochastic depth rate.
drop_path_uniform : bool
Apply uniform drop rate across blocks.
weight_init : str
Weight init scheme.
init_values : float
Layer-scale init values.
act_layer : nn.Module
MLP activation layer.
block_fn : nn.Module
Transformer block class.
ffn_layer : DinoVisionTransformerFFNLayer
Type of FFN layer to use, can be DinoVisionTransformerFFNLayer.MLP,
DinoVisionTransformerFFNLayer.SWIGLU, DinoVisionTransformerFFNLayer.SWIGLU_FUSED,
or DinoVisionTransformerFFNLayer.IDENTITY. Default is DinoVisionTransformerFFNLayer.MLP.
block_chunks : int
Split block sequence into block_chunks units for FSDP wrap.
num_register_tokens : int
Number of extra tokens for the model to deposit information (so-called "registers"). Default is 0.
interpolate_antialias : str
Flag to apply anti-aliasing when interpolating positional embeddings.
interpolate_offset : float
Work-around offset to apply when interpolating positional embeddings.
"""
super().__init__(
dim=DinoVisionTransformerDim.TWO_D,
img_size=img_size,
patch_size=patch_size,
in_chans=in_chans,
embed_dim=embed_dim,
depth=depth,
num_heads=num_heads,
mlp_ratio=mlp_ratio,
qkv_bias=qkv_bias,
ffn_bias=ffn_bias,
proj_bias=proj_bias,
drop_path_rate=drop_path_rate,
drop_path_uniform=drop_path_uniform,
init_values=init_values,
act_layer=act_layer,
block_fn=block_fn,
ffn_layer=ffn_layer,
block_chunks=block_chunks,
num_register_tokens=num_register_tokens,
interpolate_antialias=interpolate_antialias,
interpolate_offset=interpolate_offset,
)
@override
def _interpolate_and_reshape_pos_embed(
self, patch_pos_embed: torch.Tensor, patches_resolution: tuple[int, int], dim: int, interpolation_kwargs: dict
) -> torch.Tensor:
"""Interpolate and reshape 2D patch positional embeddings.
Parameters
----------
patch_pos_embed : torch.Tensor
Positional embedding tensor of shape (1, N, C).
patches_resolution : tuple of ints
Number of patches along each spatial dimension.
dim : int
Embedding dimension.
interpolation_kwargs : dict
Arguments passed to `F.interpolate`.
Returns
-------
torch.Tensor
Reshaped and interpolated tensor of shape (1, H, W, C),
where H, W are the number of patches along height and width.
"""
patch_pos_embed = patch_pos_embed.reshape(1, *patches_resolution, dim).permute(0, 3, 1, 2)
patch_pos_embed = nn.functional.interpolate(
patch_pos_embed,
mode="bicubic",
antialias=self.interpolate_antialias,
**interpolation_kwargs,
)
return patch_pos_embed.permute(0, 2, 3, 1)
def vit_small(
patch_size: int | tuple[int, int] = 16,
num_register_tokens: int = 0,
**kwargs,
) -> DinoVisionTransformer:
"""Builds a small 2d vision transformer with 384-dimensional embeddings, 12 layers, 6 heads, and 4x MLP ratio.
Parameters
----------
patch_size : int or tuple[int, int]
Patch size, either a single integer or a tuple of two integers (height, width). Default is 16.
num_register_tokens : int
Number of extra tokens for the model to deposit information (so-called "registers"). Default is 0.
kwargs : dict
Additional keyword arguments to pass to the :class:`DinoVisionTransformer` constructor.
Returns
-------
DinoVisionTransformer
A small 2d vision transformer.
"""
model = DinoVisionTransformer(
patch_size=patch_size,
embed_dim=384,
depth=12,
num_heads=6,
mlp_ratio=4,
block_fn=partial(Block, attn_class=MemEffAttention),
num_register_tokens=num_register_tokens,
**kwargs,
)
return model
def vit_base(
patch_size: int | tuple[int, int] = 16,
num_register_tokens: int = 0,
**kwargs,
) -> DinoVisionTransformer:
"""Builds a base 2d vision transformer with 768-dimensional embeddings, 12 layers, 12 heads, and 4x MLP ratio.
Parameters
----------
patch_size : int or tuple[int, int]
Patch size, either a single integer or a tuple of two integers (height, width). Default is 16.
num_register_tokens : int
Number of extra tokens for the model to deposit information (so-called "registers"). Default is 0.
kwargs : dict
Additional keyword arguments to pass to the :class:`DinoVisionTransformer` constructor.
Returns
-------
DinoVisionTransformer
A base 2d vision transformer.
"""
model = DinoVisionTransformer(
patch_size=patch_size,
embed_dim=768,
depth=12,
num_heads=12,
mlp_ratio=4,
block_fn=partial(Block, attn_class=MemEffAttention),
num_register_tokens=num_register_tokens,
**kwargs,
)
return model
def vit_large(
patch_size: int | tuple[int, int] = 16,
num_register_tokens: int = 0,
**kwargs,
) -> DinoVisionTransformer:
"""Builds a large 2d vision transformer with 1024-dimensional embeddings, 24 layers, 16 heads, and 4x MLP ratio.
Parameters
----------
patch_size : int or tuple[int, int]
Patch size, either a single integer or a tuple of two integers (height, width). Default is 16.
num_register_tokens : int
Number of extra tokens for the model to deposit information (so-called "registers"). Default is 0.
kwargs : dict
Additional keyword arguments to pass to the :class:`DinoVisionTransformer` constructor.
Returns
-------
DinoVisionTransformer
A large 2d vision transformer.
"""
model = DinoVisionTransformer(
patch_size=patch_size,
embed_dim=1024,
depth=24,
num_heads=16,
mlp_ratio=4,
block_fn=partial(Block, attn_class=MemEffAttention),
num_register_tokens=num_register_tokens,
**kwargs,
)
return model
def vit_giant2(
patch_size: int | tuple[int, int] = 16,
num_register_tokens: int = 0,
**kwargs,
) -> DinoVisionTransformer:
"""Builds a giant2 vision transformer with 1536-dimensional embeddings, 40 layers, 24 heads, and 4x MLP ratio.
Close to ViT-giant, with embed-dim 1536 and 24 heads => embed-dim per head 64
Parameters
----------
patch_size : int or tuple[int, int]
Patch size, either a single integer or a tuple of two integers (height, width). Default is 16.
num_register_tokens : int
Number of extra tokens for the model to deposit information (so-called "registers"). Default is 0.
kwargs : dict
Additional keyword arguments to pass to the :class:`DinoVisionTransformer` constructor.
Returns
-------
DinoVisionTransformer
A giant2 vision transformer.
"""
model = DinoVisionTransformer(
patch_size=patch_size,
embed_dim=1536,
depth=40,
num_heads=24,
mlp_ratio=4,
block_fn=partial(Block, attn_class=MemEffAttention),
num_register_tokens=num_register_tokens,
**kwargs,
)
return model
|