Sync explorer_app: add auto_interp, fix image loading, remove cross-dataset panel

scripts/explorer_app.py

@@ -111,6 +111,13 @@ def _load_dataset_dict(path, label):
     if os.path.exists(names_file):
         with open(names_file) as _nf:
             feat_names = {int(k): v for k, v in json.load(_nf).items()}
+    auto_interp_file = os.path.splitext(path)[0] + '_auto_interp.json'
+    auto_interp = {}
+    if os.path.exists(auto_interp_file):
+        with open(auto_interp_file) as _af:
+            auto_interp = {int(k): v for k, v in json.load(_af).items()}
+        print(f"  Loaded {len(auto_interp)} auto-interp labels from "
+              f"{os.path.basename(auto_interp_file)}")
     entry = {
         'label':             label,
         'path':              path,
@@ -139,6 +146,7 @@ def _load_dataset_dict(path, label):
         'inference_cache':   OrderedDict(),
         'names_file':        names_file,
         'feature_names':     feat_names,
+        'auto_interp_names': auto_interp,
     }
     # Load pre-computed heatmaps sidecar if present
     sidecar = os.path.splitext(path)[0] + '_heatmaps.pt'
@@ -212,7 +220,7 @@ def _apply_dataset_globals(idx):
     global umap_backup
     global _clip_scores, _clip_vocab, _clip_embeds, _clip_scores_f32, HAS_CLIP
     global _compare_datasets
-    global feature_names, _names_file
+    global feature_names, _names_file, auto_interp_names
 
     ds = _all_datasets[idx]
     image_paths        = ds['image_paths']
@@ -242,6 +250,7 @@ def _apply_dataset_globals(idx):
     _compare_datasets  = [d for i, d in enumerate(_all_datasets) if i != idx]
     feature_names      = ds['feature_names']
     _names_file        = ds['names_file']
+    auto_interp_names  = ds['auto_interp_names']
 
     # Derived arrays used by UMAP, feature list, and callbacks
     freq          = feature_frequency.numpy()
@@ -274,6 +283,15 @@ def _save_names():
     print(f"Saved {len(feature_names)} feature names to {_names_file}")
 
 
+def _display_name(feat: int) -> str:
+    """Return the label to show in tables: manual label takes priority over auto-interp."""
+    m = feature_names.get(feat)
+    if m:
+        return m
+    a = auto_interp_names.get(feat)
+    return f"[auto] {a}" if a else ""
+
+
 # Live inference has been removed — all feature display and patch exploration
 # is driven entirely by pre-computed sidecars (_heatmaps.pt, _patch_acts.pt).
 HAS_CLIP_MODEL = False
@@ -637,11 +655,20 @@ def update_feature_display(feature_idx):
     dead = "DEAD FEATURE" if freq_val == 0 else ""
 
     feat_name = feature_names.get(feat, "")
-    name_display = (
-        f'<div style="color:#1a6faf;font-style:italic;margin:2px 0 6px 0">'
-        f'&#x1F3F7;&#xFE0E; {feat_name}</div>'
-        if feat_name else ""
-    )
+    auto_name = auto_interp_names.get(feat, "")
+    if feat_name:
+        name_display = (
+            f'<div style="color:#1a6faf;font-style:italic;margin:2px 0 6px 0">'
+            f'&#x1F3F7;&#xFE0E; {feat_name}</div>'
+        )
+    elif auto_name:
+        name_display = (
+            f'<div style="color:#5a9a5a;font-style:italic;margin:2px 0 6px 0">'
+            f'&#x1F916; {auto_name}'
+            f'<span style="font-size:10px;color:#999;margin-left:6px">(auto-interp)</span></div>'
+        )
+    else:
+        name_display = ""
 
     stats_div.text = f"""
     <h2 style="margin:4px 0">Feature {feat} <span style="color:red">{dead}</span></h2>
@@ -900,7 +927,7 @@ feature_list_source = ColumnDataSource(data=dict(
     frequency=freq_np[_init_order].tolist(),
     mean_act=mean_act_np[_init_order].tolist(),
     p75_val=p75_np[_init_order].tolist(),
-    name=[feature_names.get(int(i), "") for i in _init_order],
+    name=[_display_name(int(i)) for i in _init_order],
 ))
 
 feature_table = DataTable(
@@ -954,7 +981,7 @@ def _apply_order(order):
         frequency=freq_np[order].tolist(),
         mean_act=mean_act_np[order].tolist(),
         p75_val=p75_np[order].tolist(),
-        name=[feature_names.get(int(i), "") for i in order],
+        name=[_display_name(int(i)) for i in order],
     )
 
 
@@ -970,7 +997,7 @@ def _update_table_names():
         frequency=freq_np[order].tolist(),
         mean_act=mean_act_np[order].tolist(),
         p75_val=p75_np[order].tolist(),
-        name=[feature_names.get(int(i), "") for i in order],
+        name=[_display_name(int(i)) for i in order],
     )
 
 
@@ -1361,7 +1388,7 @@ if HAS_CLIP:
             clip_score=[float(scores_vec[i]) for i in top_indices],
             frequency=[int(feature_frequency[i].item()) for i in top_indices],
             mean_act=[float(feature_mean_act[i].item()) for i in top_indices],
-            name=[feature_names.get(int(i), "") for i in top_indices],
+            name=[_display_name(int(i)) for i in top_indices],
         )
         clip_result_div.text = (
             f'<span style="color:#1a6faf"><b>{len(top_indices)}</b> features for '