File size: 8,101 Bytes
300c478 | 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 | # 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 __future__ import annotations
import math
from typing import Union, Tuple, Optional
import torch
import torch.nn.functional as F
from torch import nn
from einops import rearrange
from .radio_cls import ClsToken
input_dim_t = Union[int, Tuple[int, int]]
class ViTPatchGenerator(nn.Module):
def __init__(self, patch_size: int, embed_dim: int, input_dims: input_dim_t, abs_pos: bool=True, normalize_patches: bool=False, cls_token: bool=False, max_input_dims: Optional[input_dim_t]=None, pos_dropout: float=0.0, return_pos_enc: bool=False, num_cls_tokens: int=1, register_multiple: Optional[int]=None, num_registers: Optional[int]=None, patch_bias: bool=False, device=None, dtype=None):
super().__init__()
if isinstance(input_dims, int):
input_dims = (input_dims, input_dims)
if max_input_dims is None:
max_input_dims = input_dims
if isinstance(max_input_dims, int):
max_input_dims = (max_input_dims, max_input_dims)
max_input_dims = tuple((int(math.ceil(d / patch_size) * patch_size) for d in max_input_dims))
self.cpe_mode = max_input_dims != input_dims
self.pos_dropout = pos_dropout
self.return_pos_enc = return_pos_enc
factory = dict(device=device, dtype=dtype)
self.patch_size = patch_size
self.abs_pos = abs_pos
self.embed_dim = embed_dim
self.num_rows = max_input_dims[0] // patch_size
self.num_cols = max_input_dims[1] // patch_size
self.input_dims = tuple((d // patch_size for d in input_dims))
self.num_patches = self.num_rows * self.num_cols
self.max_input_dims = max_input_dims
self.im_to_patches = Im2Patches(patch_size)
self.embedder = ViTPatchLinear(patch_size, embed_dim, bias=patch_bias, **factory)
if abs_pos:
scale = embed_dim ** -0.5
self.pos_embed = nn.Parameter(torch.randn(1, self.num_patches, embed_dim, **factory) * scale)
self.cls_token = ClsToken(embed_dim, num_tokens=num_cls_tokens, enabled=cls_token, register_multiple=register_multiple, num_registers=num_registers)
self.patch_normalizer = nn.LayerNorm(embed_dim) if normalize_patches else nn.Identity()
self.num_video_frames = None
def forward(self, x: torch.Tensor) -> torch.Tensor:
patches = self.embed_patches(x)
(patches, pos_enc) = self.apply_pos_enc(patches, input_size=x.shape[2:])
patches = self.cls_token(patches)
patches = self.patch_normalizer(patches)
if self.return_pos_enc:
return (patches, pos_enc)
return patches
@property
def apply_cls_token(self):
return self.cls_token.enabled
@property
def num_cls_tokens(self):
return self.cls_token.num_tokens
@property
def num_cls_patches(self):
return self.cls_token.num_patches
@property
def num_registers(self):
return self.cls_token.num_registers
@property
def num_skip(self):
return self.num_cls_tokens + self.num_registers
def no_weight_decay(self):
return ['pos_embed']
def _load_embed(self, src_embed: torch.Tensor, targ_embed: nn.Parameter):
if src_embed.shape != targ_embed.shape:
src_size = int(math.sqrt(src_embed.shape[1]))
assert src_size ** 2 == src_embed.shape[1], 'Unable to interpolate non-square embedding'
src_embed = rearrange(src_embed, 'b (h w) c -> b c h w', h=src_size, w=src_size)
src_embed = F.interpolate(src_embed, size=(self.num_rows, self.num_cols), mode='bicubic', align_corners=True, antialias=False)
src_embed = rearrange(src_embed, 'b c h w -> b (h w) c')
targ_embed.data.copy_(src_embed)
def _load_projection(self, src_proj_weight: torch.Tensor, targ_proj_weight: torch.Tensor):
if src_proj_weight.shape != targ_proj_weight.shape:
src_patch_size = int(math.sqrt(src_proj_weight.shape[1] // 3))
assert src_patch_size ** 2 * 3 == src_proj_weight.shape[1], 'Unable to interpolate non-square patch size'
src_proj_weight = rearrange(src_proj_weight, 'b (c h w) -> b c h w', c=3, h=src_patch_size, w=src_patch_size)
src_proj_weight = F.interpolate(src_proj_weight, size=(self.patch_size, self.patch_size), mode='bicubic', align_corners=True, antialias=False)
src_proj_weight = rearrange(src_proj_weight, 'b c h w -> b (c h w)')
targ_proj_weight.data.copy_(src_proj_weight)
def embed_patches(self, x: torch.Tensor) -> torch.Tensor:
patches = self.im_to_patches(x)
patches = self.embedder(patches)
return patches
def apply_pos_enc(self, patches: torch.Tensor, patch_idxs: Optional[torch.Tensor]=None, input_size: Optional[Tuple[int, int]]=None) -> torch.Tensor:
if not self.abs_pos:
return patches
pos_enc = self.get_pos_enc(patches.shape[0], patch_idxs, input_size)
pos_enc_drop = pos_enc
return (patches + pos_enc_drop, pos_enc)
def get_pos_enc(self, batch_size: int, patch_idxs: Optional[torch.Tensor]=None, input_size: Optional[Tuple[int, int]]=None) -> torch.Tensor:
if input_size is None:
input_dims = self.input_dims
else:
input_dims = tuple((d // self.patch_size for d in input_size))
pos_embed = self._get_pos_embeddings(batch_size, input_dims)
if patch_idxs is None:
return pos_embed
exp_patch_idxs = patch_idxs.unsqueeze(-1).expand(-1, -1, pos_embed.shape[-1])
pos_embed = torch.gather(pos_embed.expand(patch_idxs.shape[0], -1, -1), dim=1, index=exp_patch_idxs)
return pos_embed
def _get_pos_embeddings(self, batch_size: int, input_dims: Tuple[int, int]):
if (self.num_rows, self.num_cols) == input_dims:
return self.pos_embed
pos_embed = self.pos_embed.reshape(1, self.num_rows, self.num_cols, -1).permute(0, 3, 1, 2)
def window_select(pos_embed):
if input_dims[0] < pos_embed.shape[-2]:
pos_embed = pos_embed[..., :input_dims[0], :]
if input_dims[1] < pos_embed.shape[-1]:
pos_embed = pos_embed[..., :, :input_dims[1]]
return pos_embed
if self.cpe_mode:
max_dim = max(input_dims)
pos_embed = F.interpolate(pos_embed.float(), size=(max_dim, max_dim), align_corners=False, mode='bilinear').to(pos_embed.dtype)
pos_embed = window_select(pos_embed)
else:
pos_embed = window_select(pos_embed)
if pos_embed.shape[-2:] != input_dims:
pos_embed = F.interpolate(pos_embed.float(), size=input_dims, align_corners=False, mode='bilinear').to(pos_embed.dtype)
pos_embed = pos_embed.flatten(2).permute(0, 2, 1)
return pos_embed
class Im2Patches(nn.Module):
def __init__(self, patch_size: int):
super().__init__()
self.patch_size = patch_size
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.patch_size == 1:
patches = x.flatten(2)
patches = patches.permute(0, 2, 1)
return patches
py = x.shape[-2] // self.patch_size
px = x.shape[-1] // self.patch_size
patches = rearrange(x, 'b c (py yy) (px xx) -> b (py px) (c yy xx)', py=py, yy=self.patch_size, px=px, xx=self.patch_size)
return patches
class ViTPatchLinear(nn.Linear):
def __init__(self, patch_size: int, embed_dim: int, bias: bool=False, **factory):
super().__init__(3 * patch_size ** 2, embed_dim, bias=bias, **factory)
self.patch_size = patch_size
|