File size: 7,581 Bytes
1aef2aa | 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 | import torch
import torch.nn as nn
import torch.nn.functional as F
from torch_geometric.nn import GATv2Conv, knn_graph
from torch_geometric.utils import to_dense_batch
from config import CONFIG
class GaussianGAT(nn.Module):
"""
Phase I: Local Geometric Context
A Graph Attention Layer (GATv2) that encodes the local jet structure using
spatial distance (dR^2) to bias attention weights based on spatial proximity.
"""
def __init__(self, in_channels, out_channels, heads=4, dropout=0.1):
super().__init__()
# add_self_loops=False is critical because self-loops are handled in knn_graph
self.gat = GATv2Conv(
in_channels, out_channels, heads=heads,
concat=False, add_self_loops=False, edge_dim=1
)
self.norm = nn.LayerNorm(out_channels)
self.dropout = nn.Dropout(dropout)
def forward(self, x, edge_index, dist):
identity = x
out = self.gat(x, edge_index, edge_attr=dist)
out = self.norm(out)
out = F.gelu(out)
out = self.dropout(out)
return out + identity
class GlobalTransformerBlock(nn.Module):
"""
Phase II: Global Context
A Transformer Encoder that captures long-range correlations across the event,
such as momentum balance and event-wide pileup density.
"""
def __init__(self, embed_dim, num_heads=4, num_layers=2, dropout=0.1):
super().__init__()
encoder_layer = nn.TransformerEncoderLayer(
d_model=embed_dim,
nhead=num_heads,
dim_feedforward=embed_dim * 4,
dropout=dropout,
activation='gelu',
batch_first=True,
norm_first=True
)
self.transformer = nn.TransformerEncoder(encoder_layer, num_layers=num_layers)
def forward(self, x, batch):
x_dense, mask = to_dense_batch(x, batch)
padding_mask = ~mask
x_global_dense = self.transformer(x_dense, src_key_padding_mask=padding_mask)
x_global_flat = x_global_dense[mask]
return x_global_flat
class JetAttention(nn.Module):
"""
Phase IV: Hypergraph Aggregation
A Bipartite GAT where the Jet Node dynamically queries its constituent Tracks.
"""
def __init__(self, track_dim, jet_dim, out_dim, heads=4, dropout=0.1):
super().__init__()
self.track_proj = nn.Linear(track_dim, out_dim)
self.jet_proj = nn.Linear(jet_dim, out_dim)
self.gat = GATv2Conv(
(out_dim, out_dim), out_dim, heads=heads,
concat=False, add_self_loops=False
)
self.norm = nn.LayerNorm(out_dim)
self.dropout = nn.Dropout(dropout)
def forward(self, x_track, x_jet, edge_index):
h_track = F.gelu(self.track_proj(x_track))
h_jet = F.gelu(self.jet_proj(x_jet))
h_jet_updated = self.gat((h_track, h_jet), edge_index)
h_jet_updated = self.norm(h_jet_updated)
h_jet_updated = self.dropout(h_jet_updated)
return h_jet_updated
class PhyGHT(nn.Module):
"""
PhyGHT: Physics-Guided Hypergraph Transformer
A hierarchical architecture combining local geometric graphs with global transformers
for robust pileup mitigation at the HL-LHC.
"""
def __init__(self, hidden_dim=None, num_heads=None, num_layers=None, k_neighbors=None):
super().__init__()
hidden_dim = hidden_dim or CONFIG['HIDDEN_DIM']
num_heads = num_heads or CONFIG['NUM_HEADS']
transformer_layers = num_layers or CONFIG['NUM_LAYERS']
self.k_neighbors = k_neighbors or CONFIG['K_NEIGHBORS']
dropout = CONFIG['DROPOUT']
input_track_dim = 6 # [pT, eta, phi, m, d0, z0]
input_jet_dim = 4 # [pT, eta, phi, m]
# Encoders
self.track_lin = nn.Sequential(
nn.Linear(input_track_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.GELU()
)
self.jet_lin = nn.Sequential(
nn.Linear(input_jet_dim, hidden_dim),
nn.LayerNorm(hidden_dim),
nn.GELU()
)
# Phase I: Local Geometry
self.local_gat = GaussianGAT(hidden_dim, hidden_dim, heads=num_heads, dropout=dropout)
# Phase II: Global Context
self.global_transformer = GlobalTransformerBlock(
embed_dim=hidden_dim,
num_heads=num_heads,
num_layers=transformer_layers,
dropout=dropout
)
# Phase III: Pileup Suppression Gate (PSG)
self.gate_net = nn.Sequential(
nn.Linear(hidden_dim, 64),
nn.ReLU(),
nn.Dropout(dropout),
nn.Linear(64, 1),
nn.Sigmoid()
)
# Phase IV: Hypergraph Aggregation
self.jet_attention = JetAttention(hidden_dim, hidden_dim, hidden_dim, heads=num_heads, dropout=dropout)
# Fusion & Prediction Heads
self.jet_fusion = nn.Sequential(
nn.Linear(hidden_dim * 2, hidden_dim),
nn.ReLU(),
nn.Dropout(dropout)
)
self.head_efrac = nn.Sequential(nn.Linear(hidden_dim, 64), nn.GELU(), nn.Linear(64, 1), nn.Sigmoid())
self.head_mfrac = nn.Sequential(nn.Linear(hidden_dim, 64), nn.GELU(), nn.Linear(64, 1), nn.Sigmoid())
def forward(self, data):
x = data['track'].x
batch = data['track'].batch
# 1. Spatial Geometry Encoding
eta, phi = x[:, 1], x[:, 2]
pos = torch.stack([eta, phi], dim=1)
edge_index_knn = knn_graph(pos, k=self.k_neighbors, batch=batch, loop=True)
row, col = edge_index_knn
d_eta = eta[row] - eta[col]
d_phi = phi[row] - phi[col]
dist_sq = (d_eta**2 + d_phi**2).view(-1, 1)
# 2. Hybrid Processing
h = self.track_lin(x)
h_local = self.local_gat(h, edge_index_knn, dist_sq)
h_global = self.global_transformer(h_local, batch)
h_combined = h_local + h_global
# 3. Gating & Aggregation
signal_prob = self.gate_net(h_combined)
h_weighted = h_combined * signal_prob
bipartite_edge_index = data['jet', 'contains', 'track'].edge_index
track_to_jet_index = torch.stack([bipartite_edge_index[1], bipartite_edge_index[0]], dim=0)
h_jet_raw = self.jet_lin(data['jet'].x)
h_jet_aggr = self.jet_attention(h_weighted, h_jet_raw, track_to_jet_index)
# 4. Prediction
h_jet_final = self.jet_fusion(torch.cat([h_jet_raw, h_jet_aggr], dim=-1))
pred_e = self.head_efrac(h_jet_final).squeeze(-1)
pred_m = self.head_mfrac(h_jet_final).squeeze(-1)
return pred_e, pred_m, signal_prob
if __name__ == "__main__":
from torch_geometric.data import HeteroData
print("--- Testing PhyGHT Architecture ---")
data = HeteroData()
data['jet'].x = torch.randn(5, 4)
data['track'].x = torch.randn(100, 6)
data['track'].batch = torch.zeros(100).long()
jet_idxs = torch.arange(5).repeat_interleave(10)
trk_idxs = torch.randint(0, 100, (50,))
data['jet', 'contains', 'track'].edge_index = torch.stack([jet_idxs, trk_idxs], dim=0)
model = PhyGHT(hidden_dim=32, num_heads=2, num_layers=2, k_neighbors=5)
try:
e, m, p = model(data)
print("\nForward Pass Successful!")
print(f"Energy Pred Shape: {e.shape}")
print(f"Mass Pred Shape: {m.shape}")
print(f"Signal Prob Shape: {p.shape}")
except Exception as e:
print(f"\nError: {e}") |