File size: 7,342 Bytes
930ea3d | 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 | from __future__ import annotations
import re
from pathlib import Path
from typing import Dict, List, Optional, Tuple
import numpy as np
import torch
from torch_geometric.data import Data
from src.data_builder import featurize_smiles, TargetScaler
from src.model import build_model
from src.utils import to_device, apply_inverse_transform
# -------------------------
# Unit correction (ML only)
# -------------------------
POST_SCALE = {
"td": 1e-7,
"dif": 1e-5,
"visc": 1e-3,
}
def _load_scaler_compat(path: Path) -> TargetScaler:
blob = torch.load(path, map_location="cpu")
if "mean" not in blob or "std" not in blob:
raise RuntimeError(f"Unrecognized target_scaler format: {path}")
ts = TargetScaler(
transforms=blob.get("transforms", None),
eps=blob.get("eps", None),
)
ts.load_state_dict({
"mean": blob["mean"].float(),
"std": blob["std"].float(),
"transforms": blob.get("transforms", ts.transforms),
"eps": blob.get("eps", ts.eps),
})
ts.targets = [str(t).lower() for t in blob.get("targets", [])]
return ts
def _infer_seed(path: Path) -> Optional[int]:
m = re.search(r"_([0-9]+)\.pt$", path.name)
return int(m.group(1)) if m else None
def _make_one_graph(smiles: str, T: int, fid_idx: int = 0) -> Data:
x, edge_index, edge_attr = featurize_smiles(smiles)
d = Data(
x=x,
edge_index=edge_index,
edge_attr=edge_attr,
y=torch.zeros(1, T),
y_mask=torch.zeros(1, T, dtype=torch.bool),
fid_idx=torch.tensor([fid_idx], dtype=torch.long),
)
d.smiles = smiles
return d
class MultiTaskEnsemblePredictor:
"""
Multi-task ensemble:
models/multitask_models/{task}_model_{seed}.pt
models/multitask_models/{task}_scalar_{seed}.pt
"""
def __init__(self, models_dir: str = "models/multitask_models", device: str = "cpu"):
self.models_dir = Path(models_dir)
self.device = torch.device(device if device == "cuda" and torch.cuda.is_available() else "cpu")
self._cache: Dict[Tuple[str, int], Tuple[Optional[torch.nn.Module], TargetScaler, dict]] = {}
def available_seeds(self, task: str) -> List[int]:
task = task.strip().lower()
seeds = []
for p in self.models_dir.glob(f"{task}_model_*.pt"):
s = _infer_seed(p)
if s is not None:
seeds.append(s)
return sorted(set(seeds))
def _load_one_meta(self, task: str, seed: int):
task = task.strip().lower()
key = (task, seed)
if key in self._cache:
return self._cache[key]
ckpt_path = self.models_dir / f"{task}_model_{seed}.pt"
scaler_path = self.models_dir / f"{task}_scalar_{seed}.pt"
if not ckpt_path.exists() or not scaler_path.exists():
raise FileNotFoundError(f"Missing model/scaler for task={task} seed={seed}")
ckpt = torch.load(ckpt_path, map_location=self.device)
state_dict = ckpt["model"]
train_args = ckpt.get("args", {})
scaler = _load_scaler_compat(scaler_path)
task_names = list(getattr(scaler, "targets", []))
if not task_names:
raise RuntimeError(f"No targets found in scaler: {scaler_path}")
if "fid_embed.weight" in state_dict:
num_fids = state_dict["fid_embed.weight"].shape[0]
else:
num_fids = 1
meta = {
"train_args": train_args,
"task_names": task_names,
"num_fids": num_fids,
}
self._cache[key] = (None, scaler, meta)
return self._cache[key]
def _build_if_needed(self, task: str, seed: int, in_dim_node: int, in_dim_edge: int):
task = task.strip().lower()
key = (task, seed)
model, scaler, meta = self._cache[key]
if model is not None:
return model, scaler, meta
train_args = meta["train_args"]
task_names = meta["task_names"]
num_fids = meta["num_fids"]
model = build_model(
in_dim_node=in_dim_node,
in_dim_edge=in_dim_edge,
task_names=task_names,
num_fids=num_fids,
gnn_type=train_args.get("gnn_type", "gine"),
gnn_emb_dim=train_args.get("gnn_emb_dim", 256),
gnn_layers=train_args.get("gnn_layers", 5),
gnn_norm=train_args.get("gnn_norm", "batch"),
gnn_readout=train_args.get("gnn_readout", "mean"),
gnn_act=train_args.get("gnn_act", "relu"),
gnn_dropout=train_args.get("gnn_dropout", 0.0),
gnn_residual=train_args.get("gnn_residual", True),
fid_emb_dim=train_args.get("fid_emb_dim", 64),
use_film=train_args.get("use_film", True),
use_task_embed=train_args.get("use_task_embed", True),
task_emb_dim=train_args.get("task_emb_dim", 32),
head_hidden=train_args.get("head_hidden", 512),
head_depth=train_args.get("head_depth", 2),
head_act=train_args.get("head_act", "relu"),
head_dropout=train_args.get("head_dropout", 0.0),
heteroscedastic=train_args.get("heteroscedastic", False),
fid_emb_l2=0.0,
task_emb_l2=0.0,
use_task_uncertainty=train_args.get("task_uncertainty", False),
).to(self.device)
ckpt_path = self.models_dir / f"{task}_model_{seed}.pt"
ckpt = torch.load(ckpt_path, map_location=self.device)
model.load_state_dict(ckpt["model"], strict=True)
model.eval()
self._cache[key] = (model, scaler, meta)
return model, scaler, meta
def predict_mean_std(self, smiles: str, prop_key: str, task: str) -> Tuple[Optional[float], Optional[float], Dict[int, float]]:
task = task.strip().lower()
prop_key = prop_key.lower()
seeds = self.available_seeds(task)
if not seeds:
return None, None, {}
self._load_one_meta(task, seeds[0])
_, scaler0, meta0 = self._cache[(task, seeds[0])]
targets = list(meta0["task_names"]) # already lower()
if prop_key not in targets:
return None, None, {}
t_idx = targets.index(prop_key)
T = len(targets)
try:
g = _make_one_graph(smiles, T=T, fid_idx=0)
except Exception:
return None, None, {}
in_dim_node = g.x.shape[1]
in_dim_edge = g.edge_attr.shape[1]
per_seed: Dict[int, float] = {}
with torch.no_grad():
for seed in seeds:
self._load_one_meta(task, seed)
model, scaler, meta = self._build_if_needed(task, seed, in_dim_node, in_dim_edge)
batch = to_device(g, self.device)
out = model(batch)
pred_n = out["pred"] # [1, T]
pred = apply_inverse_transform(pred_n, scaler).cpu().numpy().reshape(-1)
val = float(pred[t_idx])
# unit correction
val *= POST_SCALE.get(prop_key, 1.0)
per_seed[seed] = val
vals = np.array(list(per_seed.values()), dtype=float)
mean = float(vals.mean())
std = float(vals.std(ddof=1)) if len(vals) > 1 else 0.0
return mean, std, per_seed
|