File size: 7,620 Bytes
5c43f61 | 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 | """
SciGateFFN: Science-aware gated feed-forward network.
Learns to activate different FFN pathways based on science domain.
Uses hybrid routing: explicit domain tags preferred, fallback to learned classifier.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional, Tuple
class SciGateFFN(nn.Module):
"""
Gated FFN with science domain routing.
Learns to activate different FFN pathways for different science domains.
Gate is conditioned on detected domain (math, chemistry, biology etc).
"""
def __init__(
self,
d_model: int,
expansion: int = 4,
num_domains: int = 7,
use_domain_tags: bool = True,
):
"""
Initialize SciGateFFN.
Args:
d_model: Model dimension
expansion: FFN expansion factor (default 4)
num_domains: Number of science domains (7)
use_domain_tags: Whether to use explicit domain tags for routing
"""
super().__init__()
self.d_model = d_model
self.expansion = expansion
self.num_domains = num_domains
self.use_domain_tags = use_domain_tags
hidden_dim = d_model * expansion
# Standard SwiGLU architecture: up_proj splits into two paths
self.up_proj = nn.Linear(d_model, hidden_dim * 2, bias=False)
self.down_proj = nn.Linear(hidden_dim, d_model, bias=False)
# Domain-specific scaling factors (learnable)
# Shape: (num_domains, hidden_dim)
self.domain_gate = nn.Linear(num_domains, hidden_dim, bias=True)
# Fallback domain classifier (when tags not present)
# Simple linear classifier based on sequence representation
self.fallback_classifier = nn.Sequential(
nn.Linear(d_model, d_model // 2),
nn.SiLU(),
nn.Linear(d_model // 2, num_domains),
)
# Initialize weights
self._initialize_weights()
def _initialize_weights(self):
"""Initialize weights."""
for module in [self.up_proj, self.down_proj, self.domain_gate, self.fallback_classifier]:
if hasattr(module, 'weight'):
nn.init.normal_(module.weight, mean=0.0, std=0.02)
if hasattr(module, 'bias') and module.bias is not None:
nn.init.zeros_(module.bias)
def get_domain_one_hot(
self,
domain_ids: Optional[torch.Tensor] = None,
domain_tags: Optional[torch.Tensor] = None,
hidden_states: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Get domain one-hot vector for routing.
Hybrid strategy:
1. If domain_tags provided (explicit [MATH], [CHEM] etc), use those
2. If domain_ids provided (from data loader), use those
3. Fallback: classify from hidden_states
Args:
domain_ids: Tensor of domain IDs (batch, seq_len) or (batch,)
domain_tags: Boolean mask for domain tags (batch, seq_len, num_domains)
hidden_states: Hidden states for fallback classification (batch, seq_len, d_model)
Returns:
domain_one_hot: (batch, seq_len, num_domains)
"""
batch, seq_len, _ = hidden_states.shape if hidden_states is not None else (0, 0, 0)
if domain_tags is not None and domain_tags.any():
# Use explicit domain tags (one-hot already)
return domain_tags.float()
elif domain_ids is not None:
# Convert domain IDs to one-hot
if domain_ids.dim() == 1:
# Same domain for entire sequence
domain_one_hot = F.one_hot(domain_ids, num_classes=self.num_domains)
# Expand to sequence length
domain_one_hot = domain_one_hot.unsqueeze(1).expand(-1, seq_len, -1)
else:
# Per-token domain IDs
domain_one_hot = F.one_hot(domain_ids, num_classes=self.num_domains)
return domain_one_hot.float()
elif hidden_states is not None:
# Fallback: classify domain from hidden states
# Use mean pooling over sequence
pooled = hidden_states.mean(dim=1) # (batch, d_model)
domain_logits = self.fallback_classifier(pooled) # (batch, num_domains)
domain_probs = F.softmax(domain_logits, dim=-1)
# Expand to sequence length
return domain_probs.unsqueeze(1).expand(-1, seq_len, -1)
else:
# Uniform distribution (no domain info)
uniform = torch.ones(batch, seq_len, self.num_domains, device=hidden_states.device if hidden_states is not None else 'cpu')
return uniform / self.num_domains
def forward(
self,
x: torch.Tensor,
domain_ids: Optional[torch.Tensor] = None,
domain_tags: Optional[torch.Tensor] = None,
) -> torch.Tensor:
"""
Forward pass with domain-aware gating.
Args:
x: Input tensor (batch, seq_len, d_model)
domain_ids: Optional domain IDs (batch,) or (batch, seq_len)
domain_tags: Optional domain tag mask (batch, seq_len, num_domains)
Returns:
Output tensor (batch, seq_len, d_model)
"""
batch, seq_len, d_model = x.shape
# Get domain routing weights
domain_weights = self.get_domain_one_hot(domain_ids, domain_tags, x)
# Shape: (batch, seq_len, num_domains)
# Project to hidden dimension
up = self.up_proj(x) # (batch, seq_len, hidden_dim * 2)
up1, up2 = up.chunk(2, dim=-1) # Each: (batch, seq_len, hidden_dim)
# Apply SwiGLU activation
hidden = up1 * F.silu(up2) # (batch, seq_len, hidden_dim)
# Apply domain-specific scaling
# domain_weights: (batch, seq_len, num_domains)
# self.domain_gate.weight: (hidden_dim, num_domains) - Linear weight shape
# einsum: (batch, seq_len, num_domains) * (hidden_dim, num_domains) -> (batch, seq_len, hidden_dim)
domain_scaling = torch.einsum(
"bsd,hd->bsh",
domain_weights,
self.domain_gate.weight # (hidden_dim, num_domains)
)
# domain_scaling: (batch, seq_len, hidden_dim)
# Apply domain scaling (multiplicative gating)
hidden = hidden * domain_scaling
# Project back to model dimension
output = self.down_proj(hidden)
return output
def test_scigate_ffn():
"""Test SciGateFFN."""
batch_size = 2
seq_len = 128
d_model = 4096
num_domains = 7
ffn = SciGateFFN(d_model, expansion=4, num_domains=num_domains)
# Test with no domain info (fallback)
x = torch.randn(batch_size, seq_len, d_model)
output = ffn(x)
print(f"Input shape: {x.shape}")
print(f"Output shape: {output.shape}")
assert output.shape == x.shape
# Test with explicit domain IDs
domain_ids = torch.randint(0, num_domains, (batch_size,))
output2 = ffn(x, domain_ids=domain_ids)
assert output2.shape == x.shape
# Test with domain tags
domain_tags = torch.zeros(batch_size, seq_len, num_domains)
domain_tags[:, :, 0] = 1.0 # All math
output3 = ffn(x, domain_tags=domain_tags)
assert output3.shape == x.shape
print("SciGateFFN test passed!")
if __name__ == "__main__":
test_scigate_ffn()
|