Add metapruning/gnn.py
Browse files- metapruning/gnn.py +263 -0
metapruning/gnn.py
ADDED
|
@@ -0,0 +1,263 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
PNAConv-based Graph Metanetwork for MetaPruning.
|
| 3 |
+
|
| 4 |
+
Paper: "Meta Pruning via Graph Metanetworks" (arXiv:2506.12041)
|
| 5 |
+
Appendix B.1 describes the full architecture.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
import torch
|
| 9 |
+
import torch.nn as nn
|
| 10 |
+
import torch.nn.functional as F
|
| 11 |
+
from typing import Optional
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
class EdgeInvertor(nn.Module):
|
| 15 |
+
"""
|
| 16 |
+
EdgeInvertor from paper Eq. (16):
|
| 17 |
+
[1,1,...,1 (hidden_dim/2), -1,-1,...,-1 (hidden_dim/2)]
|
| 18 |
+
Used to encode directional edge features.
|
| 19 |
+
"""
|
| 20 |
+
def __init__(self, hidden_dim: int):
|
| 21 |
+
super().__init__()
|
| 22 |
+
half = hidden_dim // 2
|
| 23 |
+
self.register_buffer("invertor", torch.cat([
|
| 24 |
+
torch.ones(half),
|
| 25 |
+
-torch.ones(hidden_dim - half)
|
| 26 |
+
]))
|
| 27 |
+
|
| 28 |
+
def forward(self, edge_features: torch.Tensor) -> torch.Tensor:
|
| 29 |
+
return edge_features * self.invertor.unsqueeze(0)
|
| 30 |
+
|
| 31 |
+
|
| 32 |
+
class PNAConvLayer(nn.Module):
|
| 33 |
+
"""
|
| 34 |
+
A single message passing layer for the metanetwork.
|
| 35 |
+
Implements Eqs. (12-19) from Appendix B.1.
|
| 36 |
+
|
| 37 |
+
Each layer:
|
| 38 |
+
1. Computes messages m_ij and m'_ij (Eq. 14-15)
|
| 39 |
+
2. Aggregates messages using MEAN, STD, MAX, MIN (Eq. 17-18)
|
| 40 |
+
3. Updates node features (Eq. 17)
|
| 41 |
+
4. Updates edge features (Eq. 19)
|
| 42 |
+
"""
|
| 43 |
+
|
| 44 |
+
def __init__(self, hidden_dim: int, dropout: float = 0.0):
|
| 45 |
+
super().__init__()
|
| 46 |
+
self.hidden_dim = hidden_dim
|
| 47 |
+
|
| 48 |
+
# MLPs for computing messages (Eq. 14-15)
|
| 49 |
+
self.mlp_node_1 = nn.Sequential(
|
| 50 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 51 |
+
nn.SiLU(),
|
| 52 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 53 |
+
)
|
| 54 |
+
self.mlp_node_2 = nn.Sequential(
|
| 55 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 56 |
+
nn.SiLU(),
|
| 57 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 58 |
+
)
|
| 59 |
+
|
| 60 |
+
# Edge invertor
|
| 61 |
+
self.edge_invertor = EdgeInvertor(hidden_dim)
|
| 62 |
+
|
| 63 |
+
# MLP for aggregation (Eq. 18)
|
| 64 |
+
# Input: concatenation of [MEAN, STD, MAX, MIN] -> 4 * hidden_dim
|
| 65 |
+
self.mlp_aggr = nn.Sequential(
|
| 66 |
+
nn.Linear(4 * hidden_dim, hidden_dim),
|
| 67 |
+
nn.SiLU(),
|
| 68 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 69 |
+
)
|
| 70 |
+
|
| 71 |
+
# Edge update MLPs (Eq. 19)
|
| 72 |
+
self.mlp_edge_1 = nn.Sequential(
|
| 73 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 74 |
+
nn.SiLU(),
|
| 75 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 76 |
+
)
|
| 77 |
+
self.mlp_edge_2 = nn.Sequential(
|
| 78 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 79 |
+
nn.SiLU(),
|
| 80 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 81 |
+
)
|
| 82 |
+
|
| 83 |
+
self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
|
| 84 |
+
|
| 85 |
+
def forward(
|
| 86 |
+
self,
|
| 87 |
+
node_features: torch.Tensor, # [num_nodes, hidden_dim]
|
| 88 |
+
edge_index: torch.Tensor, # [2, num_edges]
|
| 89 |
+
edge_features: torch.Tensor, # [num_edges, hidden_dim]
|
| 90 |
+
) -> tuple:
|
| 91 |
+
"""
|
| 92 |
+
Returns: (updated_node_features, updated_edge_features)
|
| 93 |
+
"""
|
| 94 |
+
num_nodes = node_features.size(0)
|
| 95 |
+
num_edges = edge_index.size(1)
|
| 96 |
+
|
| 97 |
+
src = edge_index[0] # [num_edges]
|
| 98 |
+
dst = edge_index[1] # [num_edges]
|
| 99 |
+
|
| 100 |
+
# --- Node Feature Update ---
|
| 101 |
+
# Compute messages m_ij (Eq. 14): MLP^1(v_i) β MLP^2(v_j) β e_ij
|
| 102 |
+
mlp1_src = self.mlp_node_1(node_features[src]) # [num_edges, hidden_dim]
|
| 103 |
+
mlp2_dst = self.mlp_node_2(node_features[dst]) # [num_edges, hidden_dim]
|
| 104 |
+
messages_ij = mlp1_src * mlp2_dst * edge_features # Hadamard product, [num_edges, hidden_dim]
|
| 105 |
+
|
| 106 |
+
# Compute messages m'_ij (Eq. 15): MLP^1(v_j) β MLP^2(v_i) β (e_ij β EdgeInvertor)
|
| 107 |
+
mlp1_dst = self.mlp_node_1(node_features[dst])
|
| 108 |
+
mlp2_src = self.mlp_node_2(node_features[src])
|
| 109 |
+
messages_ji = mlp1_dst * mlp2_src * self.edge_invertor(edge_features)
|
| 110 |
+
|
| 111 |
+
# Aggregate messages for each node (Eq. 17-18)
|
| 112 |
+
# We use scatter operations for MEAN, STD, MAX, MIN
|
| 113 |
+
updated_nodes = node_features.clone()
|
| 114 |
+
|
| 115 |
+
for node_idx in range(num_nodes):
|
| 116 |
+
# Find edges where this node is the destination
|
| 117 |
+
mask = (dst == node_idx)
|
| 118 |
+
if mask.sum() == 0:
|
| 119 |
+
continue
|
| 120 |
+
|
| 121 |
+
msgs_to_node = messages_ij[mask] # [num_incoming, hidden_dim]
|
| 122 |
+
msgs_from_node = messages_ji[mask] # [num_outgoing, hidden_dim]
|
| 123 |
+
|
| 124 |
+
# Aggregate: MEAN, STD, MAX, MIN (Eq. 18)
|
| 125 |
+
mean_feat = msgs_to_node.mean(dim=0, keepdim=True)
|
| 126 |
+
std_feat = msgs_to_node.std(dim=0, keepdim=True) if msgs_to_node.size(0) > 1 else torch.zeros_like(mean_feat)
|
| 127 |
+
max_feat = msgs_to_node.max(dim=0, keepdim=True)[0]
|
| 128 |
+
min_feat = msgs_to_node.min(dim=0, keepdim=True)[0]
|
| 129 |
+
|
| 130 |
+
# PNA aggregation: MLP_Aggr([MEAN, STD, MAX, MIN])
|
| 131 |
+
aggr_input = torch.cat([mean_feat, std_feat, max_feat, min_feat], dim=-1) # [1, 4*hidden_dim]
|
| 132 |
+
aggr_output = self.mlp_aggr(aggr_input) # [1, hidden_dim]
|
| 133 |
+
|
| 134 |
+
# Update node (Eq. 17): v_i += PNA_Aggr(m_ij) + PNA_Aggr(m'_ij)
|
| 135 |
+
mean_feat_ji = msgs_from_node.mean(dim=0, keepdim=True)
|
| 136 |
+
std_feat_ji = msgs_from_node.std(dim=0, keepdim=True) if msgs_from_node.size(0) > 1 else torch.zeros_like(mean_feat_ji)
|
| 137 |
+
max_feat_ji = msgs_from_node.max(dim=0, keepdim=True)[0]
|
| 138 |
+
min_feat_ji = msgs_from_node.min(dim=0, keepdim=True)[0]
|
| 139 |
+
aggr_input_ji = torch.cat([mean_feat_ji, std_feat_ji, max_feat_ji, min_feat_ji], dim=-1)
|
| 140 |
+
aggr_output_ji = self.mlp_aggr(aggr_input_ji)
|
| 141 |
+
|
| 142 |
+
updated_nodes[node_idx] = node_features[node_idx] + aggr_output.squeeze(0) + aggr_output_ji.squeeze(0)
|
| 143 |
+
|
| 144 |
+
updated_nodes = self.dropout(updated_nodes)
|
| 145 |
+
|
| 146 |
+
# --- Edge Feature Update (Eq. 19) ---
|
| 147 |
+
# e_ij += MLP^1(v_i) β MLP^2(v_j) β e_ij + MLP^1(v_j) β MLP^2(v_i) β (e_ij β EdgeInvertor)
|
| 148 |
+
mlp_edge1_src = self.mlp_edge_1(node_features[src])
|
| 149 |
+
mlp_edge2_dst = self.mlp_edge_2(node_features[dst])
|
| 150 |
+
edge_update_1 = mlp_edge1_src * mlp_edge2_dst * edge_features
|
| 151 |
+
|
| 152 |
+
mlp_edge1_dst = self.mlp_edge_1(node_features[dst])
|
| 153 |
+
mlp_edge2_src = self.mlp_edge_2(node_features[src])
|
| 154 |
+
edge_update_2 = mlp_edge1_dst * mlp_edge2_src * self.edge_invertor(edge_features)
|
| 155 |
+
|
| 156 |
+
updated_edges = edge_features + edge_update_1 + edge_update_2
|
| 157 |
+
updated_edges = self.dropout(updated_edges)
|
| 158 |
+
|
| 159 |
+
return updated_nodes, updated_edges
|
| 160 |
+
|
| 161 |
+
|
| 162 |
+
class Metanetwork(nn.Module):
|
| 163 |
+
"""
|
| 164 |
+
Graph Metanetwork for MetaPruning.
|
| 165 |
+
|
| 166 |
+
Architecture (from Appendix B.1, Eqs. 12-23):
|
| 167 |
+
1. Node/Edge Encoders (MLP) -> hidden_dim
|
| 168 |
+
2. N PNAConv message passing layers
|
| 169 |
+
3. Node/Edge Decoders (MLP) -> original dimensions
|
| 170 |
+
4. Residual connection: v_out = alpha * v_pred + v_in, e_out = beta * e_pred + e_in
|
| 171 |
+
|
| 172 |
+
Paper defaults:
|
| 173 |
+
- hidden_dim: 32-64 (start with 32)
|
| 174 |
+
- num_layers: 3
|
| 175 |
+
- alpha, beta: 0.01
|
| 176 |
+
"""
|
| 177 |
+
|
| 178 |
+
def __init__(
|
| 179 |
+
self,
|
| 180 |
+
node_in_dim: int,
|
| 181 |
+
edge_in_dim: int,
|
| 182 |
+
node_out_dim: int,
|
| 183 |
+
edge_out_dim: int,
|
| 184 |
+
hidden_dim: int = 32,
|
| 185 |
+
num_layers: int = 3,
|
| 186 |
+
alpha: float = 0.01,
|
| 187 |
+
beta: float = 0.01,
|
| 188 |
+
dropout: float = 0.0,
|
| 189 |
+
):
|
| 190 |
+
super().__init__()
|
| 191 |
+
|
| 192 |
+
self.hidden_dim = hidden_dim
|
| 193 |
+
self.num_layers = num_layers
|
| 194 |
+
self.alpha = alpha
|
| 195 |
+
self.beta = beta
|
| 196 |
+
|
| 197 |
+
# Node/Edge Encoders (Eq. 12-13)
|
| 198 |
+
self.node_encoder = nn.Sequential(
|
| 199 |
+
nn.Linear(node_in_dim, hidden_dim),
|
| 200 |
+
nn.SiLU(),
|
| 201 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 202 |
+
)
|
| 203 |
+
self.edge_encoder = nn.Sequential(
|
| 204 |
+
nn.Linear(edge_in_dim, hidden_dim),
|
| 205 |
+
nn.SiLU(),
|
| 206 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 207 |
+
)
|
| 208 |
+
|
| 209 |
+
# Message passing layers
|
| 210 |
+
self.convs = nn.ModuleList([
|
| 211 |
+
PNAConvLayer(hidden_dim, dropout=dropout)
|
| 212 |
+
for _ in range(num_layers)
|
| 213 |
+
])
|
| 214 |
+
|
| 215 |
+
# Node/Edge Decoders (Eq. 20-21)
|
| 216 |
+
self.node_decoder = nn.Sequential(
|
| 217 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 218 |
+
nn.SiLU(),
|
| 219 |
+
nn.Linear(hidden_dim, node_out_dim),
|
| 220 |
+
)
|
| 221 |
+
self.edge_decoder = nn.Sequential(
|
| 222 |
+
nn.Linear(hidden_dim, hidden_dim),
|
| 223 |
+
nn.SiLU(),
|
| 224 |
+
nn.Linear(hidden_dim, edge_out_dim),
|
| 225 |
+
)
|
| 226 |
+
|
| 227 |
+
def forward(
|
| 228 |
+
self,
|
| 229 |
+
node_features: torch.Tensor,
|
| 230 |
+
edge_index: torch.Tensor,
|
| 231 |
+
edge_features: torch.Tensor,
|
| 232 |
+
) -> dict:
|
| 233 |
+
"""
|
| 234 |
+
Args:
|
| 235 |
+
node_features: [num_nodes, node_in_dim]
|
| 236 |
+
edge_index: [2, num_edges]
|
| 237 |
+
edge_features: [num_edges, edge_in_dim]
|
| 238 |
+
|
| 239 |
+
Returns:
|
| 240 |
+
dict with keys 'node_pred', 'edge_pred', 'node_out', 'edge_out'
|
| 241 |
+
"""
|
| 242 |
+
# Encode (Eq. 12-13)
|
| 243 |
+
v = self.node_encoder(node_features) # [num_nodes, hidden_dim]
|
| 244 |
+
e = self.edge_encoder(edge_features) # [num_edges, hidden_dim]
|
| 245 |
+
|
| 246 |
+
# Message passing (Eq. 17-19)
|
| 247 |
+
for conv in self.convs:
|
| 248 |
+
v, e = conv(v, edge_index, e)
|
| 249 |
+
|
| 250 |
+
# Decode (Eq. 20-21)
|
| 251 |
+
node_pred = self.node_decoder(v) # [num_nodes, node_out_dim]
|
| 252 |
+
edge_pred = self.edge_decoder(e) # [num_edges, edge_out_dim]
|
| 253 |
+
|
| 254 |
+
# Residual output (Eq. 22-23)
|
| 255 |
+
node_out = self.alpha * node_pred + node_features
|
| 256 |
+
edge_out = self.beta * edge_pred + edge_features
|
| 257 |
+
|
| 258 |
+
return {
|
| 259 |
+
'node_pred': node_pred,
|
| 260 |
+
'edge_pred': edge_pred,
|
| 261 |
+
'node_out': node_out,
|
| 262 |
+
'edge_out': edge_out,
|
| 263 |
+
}
|