File size: 9,094 Bytes
01a8278 | 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 | import torch
import torch.nn as nn
from typing import Dict, Tuple
from dataclasses import dataclass
import math
# =========================
# Config
# =========================
@dataclass
class ModelConfig:
# problem sizes
n_conditions: int = 17 # true inverse-design sidebar parameter vector
n_materials: int = 4
n_vf_categories: int = 5 # Volume fraction categories: 0.1000, 0.2000, 0.3000, 0.4000, 0.5000
n_max_layer: int = 5 # Quarter layers (max 5 for quarter-angle dataset)
# model architecture
d_model: int = 256
n_heads: int = 4
n_layers: int = 6
dropout: float = 0.0
# =========================
# Model
# =========================
def timestep_embedding(t: torch.Tensor, dim: int) -> torch.Tensor:
"""
Sinusoidal timestep embedding. t: (B,)
"""
half = dim // 2
freqs = torch.exp(-math.log(10000) * torch.arange(0, half, device=t.device) / half)
args = t.float().unsqueeze(1) * freqs.unsqueeze(0)
emb = torch.cat([torch.cos(args), torch.sin(args)], dim=1)
if dim % 2 == 1:
emb = torch.cat([emb, torch.zeros_like(emb[:, :1])], dim=1)
return emb # (B, dim)
class SelfCrossAttnBlock(nn.Module):
def __init__(self, d_model, n_heads, dropout=0.0):
super().__init__()
self.self_attn = nn.MultiheadAttention(
d_model, n_heads, dropout=dropout, batch_first=True
)
self.cross_attn = nn.MultiheadAttention(
d_model, n_heads, dropout=dropout, batch_first=True
)
self.ff = nn.Sequential(
nn.Linear(d_model, 4 * d_model),
nn.SiLU(),
nn.Linear(4 * d_model, d_model),
)
self.ln1 = nn.LayerNorm(d_model)
self.ln2 = nn.LayerNorm(d_model)
self.ln3 = nn.LayerNorm(d_model)
def forward(self, x, cond_tokens, key_padding_mask=None):
"""
x: (B, N, d) <- material + vf_category + angle tokens
cond_tokens:(B, M, d) <- condition tokens (M = n_conditions)
key_padding_mask: (B, N) optional padding mask (True = mask out, False = keep)
"""
# self-attention (within tokens)
x = self.ln1(x + self.self_attn(x, x, x, key_padding_mask=key_padding_mask)[0])
# cross-attention (tokens attend to conditions)
x = self.ln2(x + self.cross_attn(x, cond_tokens, cond_tokens)[0])
# feed-forward
x = self.ln3(x + self.ff(x))
return x
class MaterialHybridDenoiser(nn.Module):
"""
Inputs:
material_t: (B,) in [0..n_materials-1] or MASK
vf_category_t: (B,) in [0..4] volume fraction category or MASK
layer_t: (B,L) in {0,1} or MASK
angle_t: (B,L) discrete category indices [0..n_angle_categories-1] or MASK (if use_discrete_angles)
OR (B,L,1) continuous (if not use_discrete_angles)
When discrete: category n_angle_categories = dead layer, n_angle_categories+1 = MASK
cond: (B,C) continuous, C = n_conditions
t: (B,) timestep
Outputs:
material logits: (B, n_materials)
vf_category_logits: (B, 5)
angle_logits: (B,L,n_angle_categories+1) # discrete angle categories + dead (if use_discrete_angles)
OR angle: (B,L,1) # angle in radians (if not use_discrete_angles)
"""
def __init__(self, cfg: ModelConfig, mask_ids: Dict[str, int], use_discrete_angles: bool = True, n_angle_categories: int = 7):
super().__init__()
self.cfg = cfg
self.L = cfg.n_max_layer
d = cfg.d_model
self.mask_ids = mask_ids
self.use_discrete_angles = use_discrete_angles
self.n_angle_categories = n_angle_categories
# +1 to include mask token for material
self.material_emb = nn.Embedding(cfg.n_materials + 1, d)
# vf_category: 5 categories (0-4) plus mask; we allocate 6
self.vf_category_emb = nn.Embedding(cfg.n_vf_categories + 1, d)
if use_discrete_angles:
# Category n_angle_categories = dead layer, n_angle_categories+1 = mask token
self.angle_emb = nn.Embedding(n_angle_categories + 2, d)
self.layer_emb = None
else:
# layer token: {MASK, 0, 1} => 3 (only needed for continuous angles)
self.layer_emb = nn.Embedding(3, d)
self.angle_in = nn.Linear(1, d)
# Condition projection: each scalar condition coefficient gets its own Linear(1, d)
# n_conditions = 7 * degree
self.cond_proj = nn.ModuleList([
nn.Linear(1, d) for _ in range(cfg.n_conditions)
])
self.blocks = nn.ModuleList([
SelfCrossAttnBlock(d, cfg.n_heads, cfg.dropout)
for _ in range(cfg.n_layers)
])
# Positional embeddings: pos 0 = material, pos 1 = vf_category, pos 2..2+L-1 = layers
self.pos_emb = nn.Embedding(2 + cfg.n_max_layer, d)
self.t_proj = nn.Linear(d, d)
enc_layer = nn.TransformerEncoderLayer(
d_model=d,
nhead=cfg.n_heads,
dropout=cfg.dropout,
batch_first=True,
)
self.encoder = nn.TransformerEncoder(enc_layer, num_layers=cfg.n_layers)
self.ln = nn.LayerNorm(d)
self.material_head = nn.Linear(d, cfg.n_materials)
self.vf_category_head = nn.Linear(d, cfg.n_vf_categories)
if use_discrete_angles:
# n_angle_categories for angles + 1 for dead layer
self.angle_head = nn.Linear(d, n_angle_categories + 1)
self.layer_head = None
else:
self.layer_head = nn.Linear(d, 2) # alive/dead
self.angle_head = nn.Linear(d, 1)
def forward(self, material_t, vf_category_t, layer_t, angle_t, cond, t):
B, L = layer_t.shape
assert L == self.L
# Project conditions if provided as raw scalars (B, C)
if cond.dim() == 2:
cond_list = []
for i in range(cond.shape[1]):
cond_list.append(self.cond_proj[i](cond[:, i:i+1].unsqueeze(-1))) # (B, 1, d)
cond = torch.cat(cond_list, dim=1) # (B, C, d)
# global tokens as a 2-token "prefix"
g_mat = self.material_emb(material_t).unsqueeze(1) # (B,1,d)
g_vf = self.vf_category_emb(vf_category_t).unsqueeze(1) # (B,1,d)
# per-layer tokens
if self.use_discrete_angles:
layer_h = self.angle_emb(angle_t) # (B, L, d)
else:
layer_h = self.layer_emb(layer_t) + self.angle_in(angle_t) # (B,L,d)
h = torch.cat([g_mat, g_vf, layer_h], dim=1) # (B, 2+L, d)
# Add positional embeddings to entire sequence
pos_indices = torch.arange(2 + self.L, device=h.device) # (2+L,)
h = h + self.pos_emb(pos_indices).unsqueeze(0) # (B, 2+L, d)
# add timestep
t_emb = timestep_embedding(t, h.size(-1)) # (B,d)
h = h + self.t_proj(t_emb).unsqueeze(1)
# Create key padding mask to enforce dead tokens are at the end
key_padding_mask = None
if self.use_discrete_angles:
dead_category = self.n_angle_categories
is_dead = (angle_t == dead_category) # (B, L)
first_dead_pos = torch.zeros(B, dtype=torch.long, device=angle_t.device)
for b in range(B):
dead_positions = torch.where(is_dead[b])[0]
if len(dead_positions) > 0:
first_dead_pos[b] = dead_positions[0].item() + 2 # +2 for global tokens offset
else:
first_dead_pos[b] = 2 + L # No dead tokens
N = 2 + L
key_padding_mask = torch.zeros(B, N, dtype=torch.bool, device=h.device)
for b in range(B):
first_invalid = first_dead_pos[b].item()
if first_invalid < 2 + L:
key_padding_mask[b, first_invalid:] = True
key_padding_mask[b, :2] = False
for block in self.blocks:
h = block(h, cond, key_padding_mask=key_padding_mask)
h = self.ln(h)
if self.use_discrete_angles:
angle_logits = self.angle_head(h[:, 2:]) # (B, L, n_angle_categories + 1)
out = {
"material_logits": self.material_head(h[:, 0]), # (B, n_materials)
"vf_category_logits": self.vf_category_head(h[:, 1]), # (B, 5)
"angle_logits": angle_logits, # (B,L,n_angle_categories+1)
}
else:
angle_raw = self.angle_head(h[:, 2:]) # (B,L,1)
angle = torch.sigmoid(angle_raw) * (math.pi / 2) # (B,L,1) in radians
out = {
"material_logits": self.material_head(h[:, 0]), # (B, n_materials)
"vf_category_logits": self.vf_category_head(h[:, 1]), # (B, 5)
"layer_logits": self.layer_head(h[:, 2:]), # (B,L,2)
"angle": angle, # (B,L,1) in radians
}
return out
|