File size: 8,154 Bytes
3ed0796 |
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 |
# Copyright (c) 2023-2024, NVIDIA CORPORATION. All rights reserved.
#
# NVIDIA CORPORATION and its licensors retain all intellectual property
# and proprietary rights in and to this software, related documentation
# and any modifications thereto. Any use, reproduction, disclosure or
# distribution of this software and related documentation without an express
# license agreement from NVIDIA CORPORATION is strictly prohibited.
from typing import List, Optional, Set, Tuple, Union
from types import MethodType
import torch
from torch import nn
from timm.models import VisionTransformer, checkpoint_seq
from timm.models.vision_transformer import Attention, Block
from .feature_normalizer import IntermediateFeatureNormalizerBase, NullIntermediateFeatureNormalizer
from .extra_models import DinoWrapper
from .vit_patch_generator import ViTPatchGenerator
from .forward_intermediates import forward_intermediates
from .dual_hybrid_vit import HybridModel
from flash_attn import flash_attn_varlen_func
def _attn_forward_pack(self: Attention, x: torch.Tensor, cu_seqlens: torch.Tensor) -> torch.Tensor:
N, C = x.shape
qkv = self.qkv(x).reshape(N, 3, self.num_heads, self.head_dim).permute(1, 0, 2, 3)
q, k, v = qkv.unbind(0)
q, k = self.q_norm(q), self.k_norm(k)
max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
x = flash_attn_varlen_func(
q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen
).reshape(N, -1)
x = self.proj(x)
x = self.proj_drop(x)
return x
def _block_forward_pack(self: Block, x: torch.Tensor, cu_seqlens: torch.Tensor) -> torch.Tensor:
x = x + self.drop_path1(self.ls1(self.attn(self.norm1(x), cu_seqlens)))
x = x + self.drop_path2(self.ls2(self.mlp(self.norm2(x))))
return x
def _forward_cpe_pack(self: VisionTransformer, images: List[torch.Tensor]) -> torch.Tensor:
device = images[0].device
x = []
seqlens = []
for image in images:
# image: [1, c, H, W] -> x: [n_cls+h*w, D], h=H/p and w=W/p
_image = self.patch_generator(image).squeeze(0)
x.append(_image)
seqlens.append(_image.shape[0])
x = torch.cat(x, dim=0)
seqlens = torch.tensor(seqlens, device=device, dtype=torch.int)
cu_seqlens = torch.cat([
torch.tensor([0], device=device, dtype=torch.int32),
torch.cumsum(seqlens, dim=0, dtype=torch.int32)
])
if getattr(self, 'grad_checkpointing', False) and not torch.jit.is_scripting():
for block in self.blocks:
x = checkpoint_seq(block, x, cu_seqlens)
else:
for block in self.blocks:
x = block(x, cu_seqlens)
x = self.norm(x)
return x, cu_seqlens
def _forward_cpe(self: VisionTransformer, x: torch.Tensor) -> torch.Tensor:
x = self.patch_generator(x)
if getattr(self, 'grad_checkpointing', False) and not torch.jit.is_scripting():
x = checkpoint_seq(self.blocks, x)
else:
x = self.blocks(x)
x = self.norm(x)
return x
def _take_indices(
num_blocks: int,
n: Optional[Union[int, List[int], Tuple[int]]],
) -> Tuple[Set[int], int]:
if isinstance(n, int):
assert n >= 0
take_indices = {x for x in range(num_blocks - n, num_blocks)}
else:
take_indices = {num_blocks + idx if idx < 0 else idx for idx in n}
return take_indices, max(take_indices)
def _forward_intermediates_cpe(
self,
x: torch.Tensor,
norm: bool = False,
**kwargs,
) -> Union[List[torch.Tensor], Tuple[torch.Tensor, List[torch.Tensor]]]:
return forward_intermediates(
self,
patch_extractor=self.patch_generator,
num_summary_tokens=self.patch_generator.num_skip,
num_cls_tokens=self.patch_generator.num_cls_tokens,
norm=self.norm if norm else lambda y: y,
x=x,
**kwargs,
)
def _forward_cpe_dinov2(self: DinoWrapper, x: torch.Tensor) -> torch.Tensor:
y = _forward_cpe(self.inner, x)
return y[:, 0], y[:, self.num_summary_tokens:]
def _forward_intermediates_cpe_dinov2(self: DinoWrapper, *args, **kwargs):
return _forward_intermediates_cpe(self.inner, *args, **kwargs)
def _enable_cpe_for_timm_vit(model: VisionTransformer,
max_img_size: Union[int, Tuple[int, int]] = 1024,
num_cls_tokens: int = 1,
pos_dropout: float = 0.1,
register_multiple: int = Optional[None],
num_registers: int = Optional[None],
support_packing: bool = False,
):
if not isinstance(model, VisionTransformer):
raise ValueError("CPE only support for VisionTransformer models!")
patch_size = model.patch_embed.patch_size[0]
embed_dim = model.embed_dim
input_dims = model.patch_embed.img_size
normalize_patches = not isinstance(model.patch_embed.norm, nn.Identity)
cls_token = model.cls_token is not None
max_img_size = int(round(max_img_size / patch_size) * patch_size)
patch_generator = ViTPatchGenerator(
patch_size=patch_size,
embed_dim=embed_dim,
input_dims=input_dims,
normalize_patches=normalize_patches,
cls_token=cls_token,
max_input_dims=max_img_size,
pos_dropout=pos_dropout,
num_cls_tokens=num_cls_tokens,
register_multiple=register_multiple,
num_registers=num_registers,
)
model.patch_generator = patch_generator
model.patch_embed = None
model.cls_token = None
model.pos_embed = None
model.pos_drop = None
model.patch_size = patch_size
model.num_cls_tokens = num_cls_tokens
model.num_registers = patch_generator.num_registers
model.forward_features = MethodType(_forward_cpe, model)
model.forward_intermediates = MethodType(_forward_intermediates_cpe, model)
if support_packing:
model.forward_features = MethodType(_forward_cpe_pack, model)
for block in model.blocks:
block.forward = MethodType(_block_forward_pack, block)
block.attn.forward = MethodType(_attn_forward_pack, block.attn)
def _enable_cpe_for_dv2_reg_vit(model: DinoWrapper,
max_img_size: Union[int, Tuple[int, int]] = 1024,
num_cls_tokens: int = 1,
pos_dropout: float = 0.1,
register_multiple: int = Optional[None],
num_registers: int = Optional[None],
):
patch_size = model.patch_size
embed_dim = model.embed_dim
input_dims = model.inner.patch_embed.patches_resolution
normalize_patches = not isinstance(model.inner.patch_embed.norm, nn.Identity)
cls_token = True
max_img_size = int(round(max_img_size / patch_size) * patch_size)
patch_generator = ViTPatchGenerator(
patch_size=patch_size,
embed_dim=embed_dim,
input_dims=input_dims,
normalize_patches=normalize_patches,
cls_token=cls_token,
max_input_dims=max_img_size,
pos_dropout=pos_dropout,
num_cls_tokens=num_cls_tokens,
register_multiple=register_multiple,
num_registers=num_registers,
patch_bias=True,
)
inner = model.inner
inner.patch_generator = patch_generator
inner.patch_embed = None
inner.cls_token = None
inner.pos_embed = None
inner.register_tokens = None
inner.patch_size = patch_size
model.forward_features = MethodType(_forward_cpe_dinov2, model)
model.forward_intermediates = MethodType(_forward_intermediates_cpe_dinov2, model)
def enable_cpe(model: nn.Module,
*args,
**kwargs,
):
if isinstance(model, VisionTransformer):
_enable_cpe_for_timm_vit(model, *args, **kwargs)
elif isinstance(model, DinoWrapper):
_enable_cpe_for_dv2_reg_vit(model, *args, **kwargs)
elif isinstance(model, HybridModel):
_enable_cpe_for_timm_vit(model.vit, *args, **kwargs)
else:
raise ValueError(f'CPE not supported for this model type: {type(model)}')
|