File size: 12,741 Bytes
2a4c86a | 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 | import math
import torch
import torch.nn as nn
import torch.nn.functional as F
from .modeling_jit_utils import VisionRotaryEmbeddingFast, get_2d_sincos_pos_embed, RMSNorm
def modulate(x, shift, scale):
return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
class BottleneckPatchEmbed(nn.Module):
def __init__(self, img_size=224, patch_size=16, in_chans=3, pca_dim=768, embed_dim=768, bias=True):
super().__init__()
img_size = (img_size, img_size)
patch_size = (patch_size, patch_size)
num_patches = (img_size[1] // patch_size[1]) * (img_size[0] // patch_size[0])
self.img_size = img_size
self.patch_size = patch_size
self.num_patches = num_patches
self.proj1 = nn.Conv2d(in_chans, pca_dim, kernel_size=patch_size, stride=patch_size, bias=False)
self.proj2 = nn.Conv2d(pca_dim, embed_dim, kernel_size=1, stride=1, bias=bias)
def forward(self, x):
_, _, height, width = x.shape
assert height == self.img_size[0] and width == self.img_size[1], (
f"Input image size ({height}*{width}) doesn't match model ({self.img_size[0]}*{self.img_size[1]})."
)
x = self.proj2(self.proj1(x)).flatten(2).transpose(1, 2)
return x
class TimestepEmbedder(nn.Module):
def __init__(self, hidden_size, frequency_embedding_size=256):
super().__init__()
self.mlp = nn.Sequential(
nn.Linear(frequency_embedding_size, hidden_size, bias=True),
nn.SiLU(),
nn.Linear(hidden_size, hidden_size, bias=True),
)
self.frequency_embedding_size = frequency_embedding_size
@staticmethod
def timestep_embedding(t, dim, max_period=10000):
half = dim // 2
freqs = torch.exp(-math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half).to(
device=t.device
)
args = t[:, None].float() * freqs[None]
embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
if dim % 2:
embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
return embedding
def forward(self, t):
t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
t_freq = t_freq.to(dtype=self.mlp[0].weight.dtype)
t_emb = self.mlp(t_freq)
return t_emb
class LabelEmbedder(nn.Module):
def __init__(self, num_classes, hidden_size):
super().__init__()
self.embedding_table = nn.Embedding(num_classes + 1, hidden_size)
self.num_classes = num_classes
def forward(self, labels):
embeddings = self.embedding_table(labels)
return embeddings
def scaled_dot_product_attention(query, key, value, dropout_p=0.0) -> torch.Tensor:
query_len, key_len = query.size(-2), key.size(-2)
scale_factor = 1 / math.sqrt(query.size(-1))
attn_bias = torch.zeros(query.size(0), 1, query_len, key_len, dtype=query.dtype, device=query.device)
with torch.amp.autocast("cuda", enabled=False):
attn_weight = query.float() @ key.float().transpose(-2, -1) * scale_factor
attn_weight += attn_bias
attn_weight = torch.softmax(attn_weight, dim=-1)
attn_weight = torch.dropout(attn_weight, dropout_p, train=True)
out = attn_weight @ value.float()
return out.to(query.dtype)
class Attention(nn.Module):
def __init__(self, dim, num_heads=8, qkv_bias=True, qk_norm=True, attn_drop=0.0, proj_drop=0.0):
super().__init__()
self.num_heads = num_heads
head_dim = dim // num_heads
self.q_norm = RMSNorm(head_dim) if qk_norm else nn.Identity()
self.k_norm = RMSNorm(head_dim) if qk_norm else nn.Identity()
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)
def forward(self, x, rope):
batch_size, num_tokens, channels = x.shape
qkv = self.qkv(x).reshape(batch_size, num_tokens, 3, self.num_heads, channels // self.num_heads).permute(2, 0, 3, 1, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
q = self.q_norm(q)
k = self.k_norm(k)
q = rope(q)
k = rope(k)
x = scaled_dot_product_attention(q, k, v, dropout_p=self.attn_drop.p if self.training else 0.0)
x = x.transpose(1, 2).reshape(batch_size, num_tokens, channels)
x = x.to(self.proj.weight.dtype)
x = self.proj(x)
x = self.proj_drop(x)
return x
class SwiGLUFFN(nn.Module):
def __init__(self, dim: int, hidden_dim: int, drop=0.0, bias=True) -> None:
super().__init__()
hidden_dim = int(hidden_dim * 2 / 3)
self.w12 = nn.Linear(dim, 2 * hidden_dim, bias=bias)
self.w3 = nn.Linear(hidden_dim, dim, bias=bias)
self.ffn_dropout = nn.Dropout(drop)
def forward(self, x):
x12 = self.w12(x)
x1, x2 = x12.chunk(2, dim=-1)
hidden = F.silu(x1) * x2
return self.w3(self.ffn_dropout(hidden))
class FinalLayer(nn.Module):
def __init__(self, hidden_size, patch_size, out_channels):
super().__init__()
self.norm_final = RMSNorm(hidden_size)
self.linear = nn.Linear(hidden_size, patch_size * patch_size * out_channels, bias=True)
self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
def forward(self, x, c):
shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
x = modulate(self.norm_final(x), shift, scale)
x = self.linear(x)
return x
class JiTBlock(nn.Module):
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, attn_drop=0.0, proj_drop=0.0):
super().__init__()
self.norm1 = RMSNorm(hidden_size, eps=1e-6)
self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, qk_norm=True, attn_drop=attn_drop, proj_drop=proj_drop)
self.norm2 = RMSNorm(hidden_size, eps=1e-6)
mlp_hidden_dim = int(hidden_size * mlp_ratio)
self.mlp = SwiGLUFFN(hidden_size, mlp_hidden_dim, drop=proj_drop)
self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True))
def forward(self, x, c, feat_rope=None):
shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(6, dim=-1)
x = x + gate_msa.unsqueeze(1) * self.attn(modulate(self.norm1(x), shift_msa, scale_msa), rope=feat_rope)
x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
return x
class JiT(nn.Module):
def __init__(
self,
input_size=256,
patch_size=16,
in_channels=3,
hidden_size=1024,
depth=24,
num_heads=16,
mlp_ratio=4.0,
attn_drop=0.0,
proj_drop=0.0,
num_classes=1000,
bottleneck_dim=128,
in_context_len=32,
in_context_start=8,
):
super().__init__()
self.in_channels = in_channels
self.out_channels = in_channels
self.patch_size = patch_size
self.num_heads = num_heads
self.hidden_size = hidden_size
self.input_size = input_size
self.in_context_len = in_context_len
self.in_context_start = in_context_start
self.num_classes = num_classes
self.t_embedder = TimestepEmbedder(hidden_size)
self.y_embedder = LabelEmbedder(num_classes, hidden_size)
self.x_embedder = BottleneckPatchEmbed(input_size, patch_size, in_channels, bottleneck_dim, hidden_size, bias=True)
num_patches = self.x_embedder.num_patches
self.pos_embed = nn.Parameter(torch.zeros(1, num_patches, hidden_size), requires_grad=False)
if self.in_context_len > 0:
self.in_context_posemb = nn.Parameter(torch.zeros(1, self.in_context_len, hidden_size), requires_grad=True)
torch.nn.init.normal_(self.in_context_posemb, std=0.02)
half_head_dim = hidden_size // num_heads // 2
hw_seq_len = input_size // patch_size
self.feat_rope = VisionRotaryEmbeddingFast(dim=half_head_dim, pt_seq_len=hw_seq_len, num_cls_token=0)
self.feat_rope_incontext = VisionRotaryEmbeddingFast(dim=half_head_dim, pt_seq_len=hw_seq_len, num_cls_token=self.in_context_len)
self.blocks = nn.ModuleList(
[
JiTBlock(
hidden_size,
num_heads,
mlp_ratio=mlp_ratio,
attn_drop=attn_drop if (depth // 4 * 3 > i >= depth // 4) else 0.0,
proj_drop=proj_drop if (depth // 4 * 3 > i >= depth // 4) else 0.0,
)
for i in range(depth)
]
)
self.final_layer = FinalLayer(hidden_size, patch_size, self.out_channels)
self.initialize_weights()
def initialize_weights(self):
def _basic_init(module):
if isinstance(module, nn.Linear):
torch.nn.init.xavier_uniform_(module.weight)
if module.bias is not None:
nn.init.constant_(module.bias, 0)
self.apply(_basic_init)
pos_embed = get_2d_sincos_pos_embed(self.pos_embed.shape[-1], int(self.x_embedder.num_patches**0.5))
self.pos_embed.data.copy_(torch.from_numpy(pos_embed).float().unsqueeze(0))
w1 = self.x_embedder.proj1.weight.data
nn.init.xavier_uniform_(w1.view([w1.shape[0], -1]))
w2 = self.x_embedder.proj2.weight.data
nn.init.xavier_uniform_(w2.view([w2.shape[0], -1]))
nn.init.constant_(self.x_embedder.proj2.bias, 0)
nn.init.normal_(self.y_embedder.embedding_table.weight, std=0.02)
nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
for block in self.blocks:
nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
nn.init.constant_(self.final_layer.linear.weight, 0)
nn.init.constant_(self.final_layer.linear.bias, 0)
def unpatchify(self, x, patch_size):
channels = self.out_channels
height = width = int(x.shape[1] ** 0.5)
assert height * width == x.shape[1]
x = x.reshape(shape=(x.shape[0], height, width, patch_size, patch_size, channels))
x = torch.einsum("nhwpqc->nchpwq", x)
images = x.reshape(shape=(x.shape[0], channels, height * patch_size, height * patch_size))
return images
def forward(self, x, t, y):
t_emb = self.t_embedder(t)
y_emb = self.y_embedder(y)
c = t_emb + y_emb
x = self.x_embedder(x)
x += self.pos_embed
for i, block in enumerate(self.blocks):
if self.in_context_len > 0 and i == self.in_context_start:
in_context_tokens = y_emb.unsqueeze(1).repeat(1, self.in_context_len, 1)
in_context_tokens += self.in_context_posemb
x = torch.cat([in_context_tokens, x], dim=1)
x = block(x, c, self.feat_rope if i < self.in_context_start else self.feat_rope_incontext)
x = x[:, self.in_context_len :]
x = self.final_layer(x, c)
output = self.unpatchify(x, self.patch_size)
return output
def JiT_B_16(**kwargs):
return JiT(depth=12, hidden_size=768, num_heads=12, bottleneck_dim=128, in_context_len=32, in_context_start=4, patch_size=16, **kwargs)
def JiT_B_32(**kwargs):
return JiT(depth=12, hidden_size=768, num_heads=12, bottleneck_dim=128, in_context_len=32, in_context_start=4, patch_size=32, **kwargs)
def JiT_L_16(**kwargs):
return JiT(depth=24, hidden_size=1024, num_heads=16, bottleneck_dim=128, in_context_len=32, in_context_start=8, patch_size=16, **kwargs)
def JiT_L_32(**kwargs):
return JiT(depth=24, hidden_size=1024, num_heads=16, bottleneck_dim=128, in_context_len=32, in_context_start=8, patch_size=32, **kwargs)
def JiT_H_16(**kwargs):
return JiT(depth=32, hidden_size=1280, num_heads=16, bottleneck_dim=256, in_context_len=32, in_context_start=10, patch_size=16, **kwargs)
def JiT_H_32(**kwargs):
return JiT(depth=32, hidden_size=1280, num_heads=16, bottleneck_dim=256, in_context_len=32, in_context_start=10, patch_size=32, **kwargs)
JiT_models = {
"JiT-B/16": JiT_B_16,
"JiT-B/32": JiT_B_32,
"JiT-L/16": JiT_L_16,
"JiT-L/32": JiT_L_32,
"JiT-H/16": JiT_H_16,
"JiT-H/32": JiT_H_32,
}
|