File size: 8,710 Bytes
9c11409 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np
from timm.models.layers import to_2tuple
class PatchEmbed_new(nn.Module):
""" Flexible Image to Patch Embedding
"""
def __init__(
self,
patch_size=16,
in_chans=3,
embed_dim=768,
stride=16,
flatten='freq'
):
super().__init__()
self.flatten = flatten
patch_size = to_2tuple(patch_size)
stride = to_2tuple(stride)
assert flatten in ['time', 'freq']
self.patch_size = patch_size
# no padding for conv
self.proj = nn.Conv2d(in_chans, embed_dim, kernel_size=patch_size, stride=stride) # with overlapped patches
def forward(self, x):
x = self.proj(x) # (B,768,64,8)
if self.flatten == 'freq':
x = x.flatten(2).transpose(1, 2) # flatten from dim
else:
x = x.transpose(-2, -1).flatten(2).transpose(1, 2)
return x
def get_2d_sincos_pos_embed_flexible(embed_dim, grid_size, cls_token=False):
"""
grid_size: int of the grid height and width
return:
pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
"""
grid_h = np.arange(grid_size[0], dtype=np.float32)
grid_w = np.arange(grid_size[1], dtype=np.float32)
grid = np.meshgrid(grid_w, grid_h) # here w goes first
grid = np.stack(grid, axis=0)
grid = grid.reshape([2, 1, grid_size[0], grid_size[1]])
pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
if cls_token:
pos_embed = np.concatenate([np.zeros([1, embed_dim]), pos_embed], axis=0)
return pos_embed
def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
assert embed_dim % 2 == 0
# use half of dimensions to encode grid_h
emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
return emb
def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
"""
embed_dim: output dimension for each position
pos: a list of positions to be encoded: size (M,)
out: (M, D)
"""
assert embed_dim % 2 == 0
omega = np.arange(embed_dim // 2, dtype=np.float32)
omega /= embed_dim / 2.0
omega = 1.0 / 10000 ** omega # (D/2,)
pos = pos.reshape(-1) # (M,)
out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product
emb_sin = np.sin(out) # (M, D/2)
emb_cos = np.cos(out) # (M, D/2)
emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
return emb
class FixedPositionalEncoder(nn.Module):
def __init__(self, pos_embed: torch.Tensor):
super().__init__()
self.positions = pos_embed
def forward(self, x: torch.Tensor, padding_mask):
return self.positions.to(x.device)
class BlockEncoder(nn.Module):
def __init__(self, blocks, norm_layer, layer_norm_first, layerdrop, dropout):
super().__init__()
self.blocks = blocks
self.norm = norm_layer
self.layer_norm_first = layer_norm_first
self.layerdrop = layerdrop
self.dropout = nn.Dropout(dropout, inplace=True)
def forward(self, x, padding_mask, alibi_bias, alibi_scale):
if self.norm is not None and not self.layer_norm_first:
x = self.norm(x)
x = self.dropout(x)
for i, blk in enumerate(self.blocks):
if (
not self.training
or self.layerdrop == 0
or (np.random.random() > self.layerdrop)
):
ab = alibi_bias
if ab is not None and alibi_scale is not None:
scale = (
alibi_scale[i]
if alibi_scale.size(0) > 1
else alibi_scale.squeeze(0)
)
ab = ab * scale.type_as(ab)
x, _ = blk(x, padding_mask, ab)
if self.norm is not None and self.layer_norm_first:
x = self.norm(x)
return x
class AltBlock(nn.Module):
def __init__(
self,
dim,
num_heads,
mlp_ratio=4.0,
qkv_bias=False,
qk_scale=None,
drop=0.0,
attn_drop=0.0,
mlp_drop=0.0,
post_mlp_drop=0.0,
drop_path=0.0,
act_layer=nn.GELU,
norm_layer=nn.LayerNorm,
layer_norm_first=True,
ffn_targets=False,
cosine_attention=False,
):
super().__init__()
self.layer_norm_first = layer_norm_first
self.ffn_targets = ffn_targets
from timm.models.vision_transformer import DropPath, Mlp
self.norm1 = norm_layer(dim)
self.attn = AltAttention(
dim,
num_heads=num_heads,
qkv_bias=qkv_bias,
qk_scale=qk_scale,
attn_drop=attn_drop,
proj_drop=drop,
cosine_attention=cosine_attention,
)
self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
self.norm2 = norm_layer(dim)
mlp_hidden_dim = int(dim * mlp_ratio)
self.mlp = Mlp(
in_features=dim,
hidden_features=mlp_hidden_dim,
act_layer=act_layer,
drop=mlp_drop,
)
self.post_mlp_dropout = nn.Dropout(post_mlp_drop, inplace=False)
def forward(self, x, padding_mask=None, alibi_bias=None):
if self.layer_norm_first:
x = x + self.drop_path(self.attn(self.norm1(x), padding_mask, alibi_bias))
r = x = self.mlp(self.norm2(x))
t = x
x = r + self.drop_path(self.post_mlp_dropout(x))
if not self.ffn_targets:
t = x
else:
x = x + self.drop_path(self.attn(x, padding_mask, alibi_bias))
r = x = self.norm1(x)
x = self.mlp(x)
t = x
x = self.norm2(r + self.drop_path(self.post_mlp_dropout(x)))
if not self.ffn_targets:
t = x
return x, t
class AltAttention(nn.Module):
def __init__(
self,
dim,
num_heads=8,
qkv_bias=False,
qk_scale=None,
attn_drop=0.0,
proj_drop=0.0,
cosine_attention=False,
):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.scale = qk_scale or head_dim ** -0.5
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.attn_drop = nn.Dropout(attn_drop)
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
self.cosine_attention = cosine_attention
if cosine_attention:
self.logit_scale = nn.Parameter(
torch.log(10 * torch.ones((num_heads, 1, 1))), requires_grad=True
)
def forward(self, x, padding_mask=None, alibi_bias=None):
B, N, C = x.shape
qkv = (
self.qkv(x)
.reshape(B, N, 3, self.num_heads, C // self.num_heads)
.permute(2, 0, 3, 1, 4) # qkv x B x H x L x D
)
q, k, v = (
qkv[0],
qkv[1],
qkv[2],
) # make torchscript happy (cannot use tensor as tuple)
dtype = q.dtype
if self.cosine_attention:
# cosine attention
attn = F.normalize(q, dim=-1) @ F.normalize(k, dim=-1).transpose(-2, -1)
logit_scale = torch.clamp(
self.logit_scale, max=torch.log(torch.tensor(1.0 / 0.01))
).exp()
attn = attn * logit_scale
else:
q = q * self.scale
attn = q @ k.transpose(-2, -1)
if alibi_bias is not None:
attn = attn.type_as(alibi_bias)
attn[:, : alibi_bias.size(1)] += alibi_bias
if padding_mask is not None and padding_mask.any():
attn = attn.masked_fill(
padding_mask.unsqueeze(1).unsqueeze(2).to(torch.bool),
float("-inf"),
)
attn = attn.softmax(dim=-1, dtype=torch.float32).to(dtype=dtype)
attn = self.attn_drop(attn)
x = (attn @ v).transpose(1, 2) #
x = x.reshape(B, N, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
|