ESPR3SS0's picture
Add metapruning/gnn.py
daa2bf5 verified
Raw History Blame Contribute Delete
9.63 kB
"""
PNAConv-based Graph Metanetwork for MetaPruning.
Paper: "Meta Pruning via Graph Metanetworks" (arXiv:2506.12041)
Appendix B.1 describes the full architecture.
"""
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Optional
class EdgeInvertor(nn.Module):
"""
EdgeInvertor from paper Eq. (16):
[1,1,...,1 (hidden_dim/2), -1,-1,...,-1 (hidden_dim/2)]
Used to encode directional edge features.
"""
def __init__(self, hidden_dim: int):
super().__init__()
half = hidden_dim // 2
self.register_buffer("invertor", torch.cat([
torch.ones(half),
-torch.ones(hidden_dim - half)
]))
def forward(self, edge_features: torch.Tensor) -> torch.Tensor:
return edge_features * self.invertor.unsqueeze(0)
class PNAConvLayer(nn.Module):
"""
A single message passing layer for the metanetwork.
Implements Eqs. (12-19) from Appendix B.1.
Each layer:
1. Computes messages m_ij and m'_ij (Eq. 14-15)
2. Aggregates messages using MEAN, STD, MAX, MIN (Eq. 17-18)
3. Updates node features (Eq. 17)
4. Updates edge features (Eq. 19)
"""
def __init__(self, hidden_dim: int, dropout: float = 0.0):
super().__init__()
self.hidden_dim = hidden_dim
# MLPs for computing messages (Eq. 14-15)
self.mlp_node_1 = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim),
)
self.mlp_node_2 = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim),
)
# Edge invertor
self.edge_invertor = EdgeInvertor(hidden_dim)
# MLP for aggregation (Eq. 18)
# Input: concatenation of [MEAN, STD, MAX, MIN] -> 4 * hidden_dim
self.mlp_aggr = nn.Sequential(
nn.Linear(4 * hidden_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim),
)
# Edge update MLPs (Eq. 19)
self.mlp_edge_1 = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim),
)
self.mlp_edge_2 = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim),
)
self.dropout = nn.Dropout(dropout) if dropout > 0 else nn.Identity()
def forward(
self,
node_features: torch.Tensor, # [num_nodes, hidden_dim]
edge_index: torch.Tensor, # [2, num_edges]
edge_features: torch.Tensor, # [num_edges, hidden_dim]
) -> tuple:
"""
Returns: (updated_node_features, updated_edge_features)
"""
num_nodes = node_features.size(0)
num_edges = edge_index.size(1)
src = edge_index[0] # [num_edges]
dst = edge_index[1] # [num_edges]
# --- Node Feature Update ---
# Compute messages m_ij (Eq. 14): MLP^1(v_i) βŠ™ MLP^2(v_j) βŠ™ e_ij
mlp1_src = self.mlp_node_1(node_features[src]) # [num_edges, hidden_dim]
mlp2_dst = self.mlp_node_2(node_features[dst]) # [num_edges, hidden_dim]
messages_ij = mlp1_src * mlp2_dst * edge_features # Hadamard product, [num_edges, hidden_dim]
# Compute messages m'_ij (Eq. 15): MLP^1(v_j) βŠ™ MLP^2(v_i) βŠ™ (e_ij βŠ™ EdgeInvertor)
mlp1_dst = self.mlp_node_1(node_features[dst])
mlp2_src = self.mlp_node_2(node_features[src])
messages_ji = mlp1_dst * mlp2_src * self.edge_invertor(edge_features)
# Aggregate messages for each node (Eq. 17-18)
# We use scatter operations for MEAN, STD, MAX, MIN
updated_nodes = node_features.clone()
for node_idx in range(num_nodes):
# Find edges where this node is the destination
mask = (dst == node_idx)
if mask.sum() == 0:
continue
msgs_to_node = messages_ij[mask] # [num_incoming, hidden_dim]
msgs_from_node = messages_ji[mask] # [num_outgoing, hidden_dim]
# Aggregate: MEAN, STD, MAX, MIN (Eq. 18)
mean_feat = msgs_to_node.mean(dim=0, keepdim=True)
std_feat = msgs_to_node.std(dim=0, keepdim=True) if msgs_to_node.size(0) > 1 else torch.zeros_like(mean_feat)
max_feat = msgs_to_node.max(dim=0, keepdim=True)[0]
min_feat = msgs_to_node.min(dim=0, keepdim=True)[0]
# PNA aggregation: MLP_Aggr([MEAN, STD, MAX, MIN])
aggr_input = torch.cat([mean_feat, std_feat, max_feat, min_feat], dim=-1) # [1, 4*hidden_dim]
aggr_output = self.mlp_aggr(aggr_input) # [1, hidden_dim]
# Update node (Eq. 17): v_i += PNA_Aggr(m_ij) + PNA_Aggr(m'_ij)
mean_feat_ji = msgs_from_node.mean(dim=0, keepdim=True)
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)
max_feat_ji = msgs_from_node.max(dim=0, keepdim=True)[0]
min_feat_ji = msgs_from_node.min(dim=0, keepdim=True)[0]
aggr_input_ji = torch.cat([mean_feat_ji, std_feat_ji, max_feat_ji, min_feat_ji], dim=-1)
aggr_output_ji = self.mlp_aggr(aggr_input_ji)
updated_nodes[node_idx] = node_features[node_idx] + aggr_output.squeeze(0) + aggr_output_ji.squeeze(0)
updated_nodes = self.dropout(updated_nodes)
# --- Edge Feature Update (Eq. 19) ---
# e_ij += MLP^1(v_i) βŠ™ MLP^2(v_j) βŠ™ e_ij + MLP^1(v_j) βŠ™ MLP^2(v_i) βŠ™ (e_ij βŠ™ EdgeInvertor)
mlp_edge1_src = self.mlp_edge_1(node_features[src])
mlp_edge2_dst = self.mlp_edge_2(node_features[dst])
edge_update_1 = mlp_edge1_src * mlp_edge2_dst * edge_features
mlp_edge1_dst = self.mlp_edge_1(node_features[dst])
mlp_edge2_src = self.mlp_edge_2(node_features[src])
edge_update_2 = mlp_edge1_dst * mlp_edge2_src * self.edge_invertor(edge_features)
updated_edges = edge_features + edge_update_1 + edge_update_2
updated_edges = self.dropout(updated_edges)
return updated_nodes, updated_edges
class Metanetwork(nn.Module):
"""
Graph Metanetwork for MetaPruning.
Architecture (from Appendix B.1, Eqs. 12-23):
1. Node/Edge Encoders (MLP) -> hidden_dim
2. N PNAConv message passing layers
3. Node/Edge Decoders (MLP) -> original dimensions
4. Residual connection: v_out = alpha * v_pred + v_in, e_out = beta * e_pred + e_in
Paper defaults:
- hidden_dim: 32-64 (start with 32)
- num_layers: 3
- alpha, beta: 0.01
"""
def __init__(
self,
node_in_dim: int,
edge_in_dim: int,
node_out_dim: int,
edge_out_dim: int,
hidden_dim: int = 32,
num_layers: int = 3,
alpha: float = 0.01,
beta: float = 0.01,
dropout: float = 0.0,
):
super().__init__()
self.hidden_dim = hidden_dim
self.num_layers = num_layers
self.alpha = alpha
self.beta = beta
# Node/Edge Encoders (Eq. 12-13)
self.node_encoder = nn.Sequential(
nn.Linear(node_in_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim),
)
self.edge_encoder = nn.Sequential(
nn.Linear(edge_in_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, hidden_dim),
)
# Message passing layers
self.convs = nn.ModuleList([
PNAConvLayer(hidden_dim, dropout=dropout)
for _ in range(num_layers)
])
# Node/Edge Decoders (Eq. 20-21)
self.node_decoder = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, node_out_dim),
)
self.edge_decoder = nn.Sequential(
nn.Linear(hidden_dim, hidden_dim),
nn.SiLU(),
nn.Linear(hidden_dim, edge_out_dim),
)
def forward(
self,
node_features: torch.Tensor,
edge_index: torch.Tensor,
edge_features: torch.Tensor,
) -> dict:
"""
Args:
node_features: [num_nodes, node_in_dim]
edge_index: [2, num_edges]
edge_features: [num_edges, edge_in_dim]
Returns:
dict with keys 'node_pred', 'edge_pred', 'node_out', 'edge_out'
"""
# Encode (Eq. 12-13)
v = self.node_encoder(node_features) # [num_nodes, hidden_dim]
e = self.edge_encoder(edge_features) # [num_edges, hidden_dim]
# Message passing (Eq. 17-19)
for conv in self.convs:
v, e = conv(v, edge_index, e)
# Decode (Eq. 20-21)
node_pred = self.node_decoder(v) # [num_nodes, node_out_dim]
edge_pred = self.edge_decoder(e) # [num_edges, edge_out_dim]
# Residual output (Eq. 22-23)
node_out = self.alpha * node_pred + node_features
edge_out = self.beta * edge_pred + edge_features
return {
'node_pred': node_pred,
'edge_pred': edge_pred,
'node_out': node_out,
'edge_out': edge_out,
}