update models

inference.py

@@ -113,7 +113,8 @@ MODEL_ALIAS = {
     "XGB": "xgb",
     "XGB_REG": "xgb_reg", 
     "POOLED": "pooled", 
-    "UNPOOLED": "unpooled"
+    "UNPOOLED": "unpooled",
+    "TRANSFORMER_WT_LOG": "transformer_wt_log",
 }
 def canon_model(label: Optional[str]) -> Optional[str]:
     if label is None:
@@ -719,15 +720,25 @@ class PeptiVersePredictor:
                     self.models[(prop_key, mode)] = obj
                 else:
                     # rebuild NN architecture
-                    self.models[(prop_key, mode)] = build_torch_model_from_ckpt(m, obj, self.device)
+                    arch = m
+                    if arch.startswith("transformer"):
+                        arch = "transformer"
+                    elif arch.startswith("mlp"):
+                        arch = "mlp"
+                    elif arch.startswith("cnn"):
+                        arch = "cnn"
+
+                    self.models[(prop_key, mode)] = build_torch_model_from_ckpt(arch, obj, self.device)
 
                 self.meta[(prop_key, mode)] = {
-                    "task_type": row.task_type,
-                    "threshold": thr,
-                    "artifact": str(art),
-                    "model_name": m,
-                    "kind": kind,
-                }
+                        "task_type": row.task_type,
+                        "threshold": thr,
+                        "artifact": str(art),
+                        "model_name": m, 
+                        "arch_name": arch,
+                        "kind": kind,
+                    }
+
 
     def _get_features_for_model(self, prop_key: str, mode: str, input_str: str):
         """
@@ -783,6 +794,14 @@ class PeptiVersePredictor:
             X, M = self._get_features_for_model(prop_key, mode, input_str)
             with torch.no_grad():
                 y = model(X, M).squeeze().float().cpu().item()
+            # invert log1p(hours) ONLY for WT half-life log models
+            model_name = meta.get("model_name", "")
+            if (
+                prop_key == "halflife"
+                and mode == "wt"
+                and model_name in {"xgb_wt_log", "transformer_wt_log"}
+            ):
+                y = float(np.expm1(y))
             if task_type == "classifier":
                 prob = float(1.0 / (1.0 + np.exp(-y)))  # sigmoid(logit)
                 out = {"property": prop_key, "mode": mode, "score": prob}
@@ -793,15 +812,22 @@ class PeptiVersePredictor:
             else:
                 return {"property": prop_key, "mode": mode, "score": float(y)}
 
-        # xgb path
         if kind == "xgb":
-            feats = self._get_features_for_model(prop_key, mode, input_str)  # (1,H)
+            feats = self._get_features_for_model(prop_key, mode, input_str)
             dmat = xgb.DMatrix(feats)
             pred = float(model.predict(dmat)[0])
+
+            # invert log1p(hours) ONLY for WT half-life log models
+            model_name = meta.get("model_name", "")
+            if (
+                prop_key == "halflife"
+                and mode == "wt"
+                and model_name in {"xgb_wt_log", "transformer_wt_log"}
+            ):
+                pred = float(np.expm1(pred))
+
             out = {"property": prop_key, "mode": mode, "score": pred}
-            if task_type == "classifier" and thr is not None:
-                out["label"] = int(pred >= float(thr))
-                out["threshold"] = float(thr)
+
             return out
 
         # joblib path (svm/enet/svr)