PhyGHT-mu60 / code /upstream_dataset.py
Rakib's picture
Publish validated, newly trained PhyGHT mu60 release (200 epochs)
1aef2aa verified
Raw History Blame Contribute Delete
4.4 kB
import os
import pickle
import torch
import numpy as np
import awkward as ak
from torch.utils.data import Dataset
from torch_geometric.data import HeteroData
from torch_geometric.loader import DataLoader
from config import CONFIG
class PhyGHTDataset(Dataset):
"""
PyTorch Geometric Dataset for PhyGHT.
Constructs a heterogeneous graph per event:
- Nodes: 'track' (6 features, 1 label) and 'jet' (4 features, 2 targets)
- Edges: ('jet', 'contains', 'track') representing constituency.
"""
def __init__(self, split='train'):
self.split = split
self.data_dir = CONFIG['PROCESSED_DATA_DIR']
self.dataset_name = CONFIG['DATASET_NAME']
self.seed = CONFIG['SEED']
file_name = f"{self.dataset_name}_{self.split}_seed_{self.seed}.pkl"
file_path = os.path.join(self.data_dir, file_name)
if not os.path.exists(file_path):
raise FileNotFoundError(f"Processed data file not found at: {file_path}")
print(f"Loading {split} dataset from {file_path}...")
with open(file_path, "rb") as f:
self.jets_ak, self.jet_trk_ids_ak, self.tracks_ak = pickle.load(f)
self.num_events = len(self.jets_ak)
print(f"Successfully loaded {self.num_events} events.")
def __len__(self):
return self.num_events
def __getitem__(self, idx):
ev_jets = self.jets_ak[idx]
ev_jet_trk_ids = self.jet_trk_ids_ak[idx]
ev_tracks = self.tracks_ak[idx]
data = HeteroData()
# Process Tracks
trk_np = ak.to_numpy(ev_tracks)
id_to_idx_map = {}
if len(trk_np) == 0:
data['track'].x = torch.zeros((0, 6), dtype=torch.float32)
data['track'].label = torch.zeros((0,), dtype=torch.float32)
else:
data['track'].x = torch.from_numpy(trk_np[:, :6].astype(np.float32))
raw_labels = trk_np[:, -1]
data['track'].label = torch.from_numpy((raw_labels == -1).astype(np.float32))
track_ids_raw = trk_np[:, 6].astype(int)
id_to_idx_map = {tid: i for i, tid in enumerate(track_ids_raw)}
# Process Jets
jet_np = ak.to_numpy(ev_jets)
if len(jet_np) == 0:
data['jet'].x = torch.zeros((0, 4), dtype=torch.float32)
data['jet'].y = torch.zeros((0, 2), dtype=torch.float32)
else:
data['jet'].x = torch.from_numpy(jet_np[:, :4].astype(np.float32))
data['jet'].y = torch.from_numpy(jet_np[:, -2:].astype(np.float32))
# Build Bipartite Edges
sources, targets = [], []
if len(jet_np) > 0 and len(trk_np) > 0:
jet_constituents_list = ev_jet_trk_ids.tolist()
for j_idx, constituents in enumerate(jet_constituents_list):
for trk_id in constituents:
if trk_id in id_to_idx_map:
t_idx = id_to_idx_map[trk_id]
sources.append(j_idx)
targets.append(t_idx)
if len(sources) > 0:
edge_index = torch.tensor([sources, targets], dtype=torch.long)
else:
edge_index = torch.empty((2, 0), dtype=torch.long)
data['jet', 'contains', 'track'].edge_index = edge_index
return data
if __name__ == "__main__":
print("--- PhyGHT Dataset Test ---")
try:
dataset = PhyGHTDataset(split='train')
except Exception as e:
print(f"Failed to initialize: {e}")
exit()
loader = DataLoader(dataset, batch_size=4, shuffle=True)
print(f"Dataset Size: {len(dataset)}")
for batch in loader:
print("\nBatch Loaded (HeteroData Batch):")
print(batch)
print(f" Track Features: {batch['track'].x.shape}")
print(f" Jet Features: {batch['jet'].x.shape}")
print(f" Track Labels: {batch['track'].label.shape}")
print(f" Edge Index: {batch['jet', 'contains', 'track'].edge_index.shape}")
if hasattr(batch['track'], 'batch'):
print(f" Track Batch Vec: {batch['track'].batch.shape}")
print(f" Max Event Idx: {batch['track'].batch.max().item()} (Should be 3 for BS=4)")
break
print("\nTest passed successfully.")