Fix HF model loading for meta tensor

app.py

@@ -303,11 +303,11 @@ def merchant_product_scores(metric: str) -> List[List[Any]]:
     for product in PRODUCTS:
         try:
             result = _predict_product(product["name"])
-        except Exception:
-            continue
+        except Exception as exc:
+            return [["ERROR", "Model loading failed", metric, type(exc).__name__, 0.0, 0.0, 0.0]]
         d = result.get("overall", {}) if metric == "Overall" else result.get("aspect_details", {}).get(metric, {})
         rows.append([product["name"], product["category"], metric, d.get("label", "Unknown"), round(_safe_float(d.get("confidence")), 4), product["price"], product["average_rating"]])
-    return rows
+    return rows or [["No rows", "Try Refresh", metric, "-", 0.0, 0.0, 0.0]]
 
 
 def _metadata_risks(features: str, categories: str, price: Any, rating: Any, count: Any) -> Dict[str, str]:
@@ -453,7 +453,7 @@ def build_app() -> gr.Blocks:
                     with gr.Column(scale=6):
                         gr.Markdown("### Product Score Monitor")
                         merchant_metric = gr.Dropdown(["Overall"] + ASPECTS, value="Overall", label="Choose overall or aspect")
-                        merchant_scores = gr.Dataframe(headers=["Product", "Category", "Metric", "Prediction", "Confidence", "Price", "Rating"], datatype=["str", "str", "str", "str", "number", "number", "number"], interactive=False)
+                        merchant_scores = gr.Dataframe(headers=["Product", "Category", "Metric", "Prediction", "Confidence", "Price", "Rating"], value=[["Click Refresh Product Scores", "", "", "", 0.0, 0.0, 0.0]], datatype=["str", "str", "str", "str", "number", "number", "number"], interactive=False)
                         refresh_scores = gr.Button("Refresh Product Scores", variant="primary")
                 gr.Markdown("### New Product Metadata Risk Screening")
                 with gr.Row():
@@ -466,7 +466,7 @@ def build_app() -> gr.Blocks:
                     new_focus = gr.Dropdown(["All"] + ASPECTS, value="All", label="Focus aspect")
                 screen_btn = gr.Button("Predict Metadata Risk", variant="primary")
                 risk_summary = gr.HTML()
-                risk_table = gr.Dataframe(headers=["Aspect", "Risk Level", "Model Signal", "Confidence", "Reason"], datatype=["str", "str", "str", "number", "str"], interactive=False)
+                risk_table = gr.Dataframe(headers=["Aspect", "Risk Level", "Model Signal", "Confidence", "Reason"], value=[["Click Predict Metadata Risk", "", "", 0.0, ""]], datatype=["str", "str", "str", "number", "str"], interactive=False)
                 gr.Markdown("### External Review Prediction")
                 external_review = gr.Textbox("The fabric is soft and the color looks good, but it runs small and the zipper feels weak.", label="External customer review", lines=4)
                 with gr.Row():
@@ -480,7 +480,7 @@ def build_app() -> gr.Blocks:
                 external_btn = gr.Button("Analyze External Review", variant="primary")
                 ext_overall = gr.HTML()
                 ext_aspects = gr.HTML()
-                ext_table = gr.Dataframe(headers=["Aspect", "Prediction", "Confidence", "Key review evidence"], datatype=["str", "str", "number", "str"], interactive=False)
+                ext_table = gr.Dataframe(headers=["Aspect", "Prediction", "Confidence", "Key review evidence"], value=[["Click Analyze External Review", "", 0.0, ""]], datatype=["str", "str", "number", "str"], interactive=False)
                 refresh_scores.click(merchant_product_scores, merchant_metric, merchant_scores)
                 merchant_metric.change(merchant_product_scores, merchant_metric, merchant_scores)
                 screen_btn.click(screen_new_product, [new_features, new_categories, new_price, new_rating, new_count, new_focus], [risk_summary, risk_table])
@@ -507,3 +507,4 @@ demo = build_app()
 
 if __name__ == "__main__":
     demo.launch()
+

src/evaluator.py

@@ -1,4 +1,4 @@
-"""Evaluation: per-aspect metrics, aggregated overall comparison.
+"""Evaluation: per-aspect metrics, aggregated overall comparison.
 
 Two label encodings live in this project (keep them straight!):
   Per-aspect (ACSA):   0=Not_Mentioned, 1=Positive, 2=Negative
@@ -41,13 +41,31 @@ OVERALL_POSITIVE = 2
 
 
 # ---------------------------------------------------------------------------
-# Loaders (all use weights_only=False 鈥?these are our own checkpoints)
+# Loaders (all use weights_only=False 閳?these are our own checkpoints)
 # ---------------------------------------------------------------------------
 
 def _load_ckpt(path: Path, device):
     return torch.load(path, map_location=device, weights_only=False)
 
 
+def _move_model_to_device(model, device):
+    """Move safely when Transformers creates meta tensors in Spaces."""
+    try:
+        return model.to(device)
+    except NotImplementedError as exc:
+        if "meta tensor" not in str(exc):
+            raise
+        logger.warning("Model contains meta tensors; using to_empty() before loading checkpoint weights.")
+        return model.to_empty(device=device)
+
+
+def _load_model_state(model, state_dict, strict=True):
+    try:
+        return model.load_state_dict(state_dict, strict=strict, assign=True)
+    except TypeError:
+        return model.load_state_dict(state_dict, strict=strict)
+
+
 def load_meta_acsa(checkpoint_dir: Path = None, meta_encoder: Optional[MetaEncoder] = None,
                    device=None):
     if checkpoint_dir is None:
@@ -71,10 +89,11 @@ def load_meta_acsa(checkpoint_dir: Path = None, meta_encoder: Optional[MetaEncod
         model = GatedAspectSemanticMetaFusionACSAModel(
             bert_name=bert_name,
             meta_in_dim=meta_in_dim,
-        ).to(device)
+        )
     else:
-        model = BertMetaFusionACSAModel(bert_name=bert_name, meta_in_dim=meta_in_dim).to(device)
-    model.load_state_dict(ckpt["model_state_dict"], strict=False)
+        model = BertMetaFusionACSAModel(bert_name=bert_name, meta_in_dim=meta_in_dim)
+    model = _move_model_to_device(model, device)
+    _load_model_state(model, ckpt["model_state_dict"], strict=False)
     model.eval()
     tokenizer = AutoTokenizer.from_pretrained(checkpoint_dir / "tokenizer")
     return model, tokenizer, meta_encoder, device
@@ -88,8 +107,9 @@ def load_acsa(checkpoint_dir: Path = None, device=None):
         device = get_device()
     ckpt = _load_ckpt(checkpoint_dir / "best.pt", device)
     bert_name = ckpt.get("config", {}).get("bert_name", cfg.BERT_MODEL_NAME)
-    model = BertACSAModel(bert_name=bert_name).to(device)
-    model.load_state_dict(ckpt["model_state_dict"])
+    model = BertACSAModel(bert_name=bert_name)
+    model = _move_model_to_device(model, device)
+    _load_model_state(model, ckpt["model_state_dict"])
     model.eval()
     tokenizer = AutoTokenizer.from_pretrained(checkpoint_dir / "tokenizer")
     return model, tokenizer, device
@@ -103,8 +123,9 @@ def load_bert_overall(checkpoint_dir: Path = None, device=None):
         device = get_device()
     ckpt = _load_ckpt(checkpoint_dir / "best.pt", device)
     bert_name = ckpt.get("config", {}).get("bert_name", cfg.BERT_MODEL_NAME)
-    model = BertOverallModel(bert_name=bert_name).to(device)
-    model.load_state_dict(ckpt["model_state_dict"])
+    model = BertOverallModel(bert_name=bert_name)
+    model = _move_model_to_device(model, device)
+    _load_model_state(model, ckpt["model_state_dict"])
     model.eval()
     tokenizer = AutoTokenizer.from_pretrained(checkpoint_dir / "tokenizer")
     return model, tokenizer, device
@@ -228,17 +249,17 @@ def overall_metrics(y_true, y_pred):
 # ---------------------------------------------------------------------------
 
 def aspect_to_overall_sentiment(aspect_preds) -> np.ndarray:
-    """Improved voting rule for aggregating per-aspect 鈫?overall.
+    """Improved voting rule for aggregating per-aspect 閳?overall.
 
     Rules (applied per review):
       1. Count n_pos and n_neg among MENTIONED aspects (skip Not_Mentioned).
-      2. If no aspect is mentioned at all 鈫?NEUTRAL (conservative default).
-      3. If n_neg 鈮?2 鈫?NEGATIVE (multiple negative aspects = clearly unhappy).
-      4. If n_neg == 1 and n_pos == 0 鈫?NEGATIVE.
-      5. If n_pos 鈮?2 and n_neg == 0 鈫?POSITIVE.
-      6. If n_pos > n_neg (but not all positive) 鈫?POSITIVE.
-      7. If n_neg > n_pos 鈫?NEGATIVE.
-      8. Otherwise (tied, or only 1 pos and 0 neg) 鈫?NEUTRAL.
+      2. If no aspect is mentioned at all 閳?NEUTRAL (conservative default).
+      3. If n_neg 閳?2 閳?NEGATIVE (multiple negative aspects = clearly unhappy).
+      4. If n_neg == 1 and n_pos == 0 閳?NEGATIVE.
+      5. If n_pos 閳?2 and n_neg == 0 閳?POSITIVE.
+      6. If n_pos > n_neg (but not all positive) 閳?POSITIVE.
+      7. If n_neg > n_pos 閳?NEGATIVE.
+      8. Otherwise (tied, or only 1 pos and 0 neg) 閳?NEUTRAL.
 
     Input  (per-aspect):  0=NM, 1=Pos, 2=Neg
     Output (overall):     0=Neg, 1=Neu, 2=Pos
@@ -253,16 +274,16 @@ def aspect_to_overall_sentiment(aspect_preds) -> np.ndarray:
 
     out = np.full(preds.shape[0], OVERALL_NEUTRAL, dtype=np.int64)
 
-    # No aspects mentioned 鈫?Neutral
-    # n_neg 鈮?2 鈫?Negative (strong signal)
+    # No aspects mentioned 閳?Neutral
+    # n_neg 閳?2 閳?Negative (strong signal)
     out[n_neg >= 2] = OVERALL_NEGATIVE
-    # n_neg == 1, n_pos == 0 鈫?Negative
+    # n_neg == 1, n_pos == 0 閳?Negative
     out[(n_neg == 1) & (n_pos == 0)] = OVERALL_NEGATIVE
-    # n_pos 鈮?2, n_neg == 0 鈫?Positive (strong signal)
+    # n_pos 閳?2, n_neg == 0 閳?Positive (strong signal)
     out[(n_pos >= 2) & (n_neg == 0)] = OVERALL_POSITIVE
-    # n_pos > n_neg and n_pos 鈮?2 鈫?Positive
+    # n_pos > n_neg and n_pos 閳?2 閳?Positive
     out[(n_pos > n_neg) & (n_pos >= 2)] = OVERALL_POSITIVE
-    # n_neg > n_pos 鈫?Negative (override the 鈮? positive if more negatives)
+    # n_neg > n_pos 閳?Negative (override the 閳? positive if more negatives)
     out[n_neg > n_pos] = OVERALL_NEGATIVE
 
     return out
@@ -324,3 +345,5 @@ def format_aspect_summary(per_aspect_pred: np.ndarray) -> Dict[str, str]:
             for i, aspect in enumerate(cfg.ASPECTS)}
 
 
+
+