Fix: saliency CPU timeout (cap 5 models), 17→20 sites, architecture 4→20 models, stale text

app.py

@@ -243,12 +243,23 @@ def get_models(atlas: str = "cc200"):
 # ── gradient saliency ──────────────────────────────────────────────────────
 
 def _compute_saliency(bw_t, adj_t, models):
+    # Cap at 5 models to keep CPU inference fast
+    sample = models[:5] if len(models) > 5 else models
     maps = []
-    for _, task in models:
-        adj = adj_t.clone().requires_grad_(True)
-        logits = task.model(bw_t, adj)
-        torch.softmax(logits, -1)[0, 1].backward()
-        maps.append(adj.grad[0].abs().detach().numpy())
+    for _, task in sample:
+        try:
+            adj = adj_t.clone().requires_grad_(True)
+            logits = task.model(bw_t, adj)
+            if isinstance(logits, tuple):
+                logits = logits[0]
+            torch.softmax(logits, -1)[0, 1].backward()
+            if adj.grad is not None:
+                maps.append(adj.grad[0].abs().detach().numpy())
+        except Exception:
+            continue
+    if not maps:
+        n = adj_t.shape[-1]
+        return np.zeros((n, n), dtype=np.float32)
     sal = np.mean(maps, axis=0)
     return (sal + sal.T) / 2
 
@@ -681,7 +692,7 @@ HEADER = """
         <div style="color:#5e6675;font-size:0.62rem;text-transform:uppercase;letter-spacing:1px;margin-top:2px">Held-out subjects</div>
       </div>
       <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:10px 18px;text-align:center">
-        <div style="font-size:1.35rem;font-weight:700;color:#f4f4f5;font-variant-numeric:tabular-nums">17</div>
+        <div style="font-size:1.35rem;font-weight:700;color:#f4f4f5;font-variant-numeric:tabular-nums">20</div>
         <div style="color:#5e6675;font-size:0.62rem;text-transform:uppercase;letter-spacing:1px;margin-top:2px">Scanner sites</div>
       </div>
       <div style="background:#161922;border:1px solid #f59e0b33;border-radius:8px;padding:10px 18px;text-align:center">
@@ -859,7 +870,7 @@ ARCHITECTURE = """
       <div style="display:flex;align-items:center;gap:8px;margin-bottom:8px">
         <span style="background:#fb923c22;color:#fb923c;font-size:0.68rem;font-weight:700;padding:2px 7px;border-radius:4px;text-transform:uppercase;letter-spacing:0.8px">Brain Modes</span>
       </div>
-      <div style="color:#cbd5e1;font-size:0.84rem;line-height:1.55">K=16 learnable directions compress the 200×200 FC matrix into 152 bilinear features — each mode specialises to a functional network (DMN, salience, FPN).</div>
+      <div style="color:#cbd5e1;font-size:0.84rem;line-height:1.55">K=16 learnable directions compress the 200×200 FC matrix into 152 bilinear features — each mode specialises to a functional network (DMN, salience, FPN). Upgraded model uses K=32 trained on AMD MI300X.</div>
     </div>
     <div style="background:#161922;border:1px solid #252a35;border-radius:8px;padding:16px 18px">
       <div style="display:flex;align-items:center;gap:8px;margin-bottom:8px">
@@ -871,7 +882,7 @@ ARCHITECTURE = """
       <div style="display:flex;align-items:center;gap:8px;margin-bottom:8px">
         <span style="background:#ef444422;color:#ef4444;font-size:0.68rem;font-weight:700;padding:2px 7px;border-radius:4px;text-transform:uppercase;letter-spacing:0.8px">LOSO</span>
       </div>
-      <div style="color:#cbd5e1;font-size:0.84rem;line-height:1.55">4 models, each trained blind to one scanner site. At inference all 4 vote — if 3/4 agree across different hardware, it's a biology signal, not an artifact.</div>
+      <div style="color:#cbd5e1;font-size:0.84rem;line-height:1.55">20 models, each trained blind to one scanner site. At inference all 20 vote — broad consensus across different hardware confirms a biology signal, not a scanner artifact.</div>
     </div>
   </div>
 
@@ -880,7 +891,7 @@ ARCHITECTURE = """
     <table style="width:100%;border-collapse:collapse;font-size:0.85rem">
       <tr><td style="padding:10px 16px;color:#8b95a7;width:150px;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Dataset</td><td style="padding:10px 16px;color:#cbd5e1">ABIDE I · 1,102 subjects · 20 acquisition sites</td></tr>
       <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Parcellation</td><td style="padding:10px 16px;color:#cbd5e1">CC200 (200 ROIs) · AAL-116 (116 ROIs) · Harvard-Oxford (111 ROIs)</td></tr>
-      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Model</td><td style="padding:10px 16px;color:#cbd5e1">AdversarialBrainModeNetwork · K=16 modes · hidden_dim=64</td></tr>
+      <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Model</td><td style="padding:10px 16px;color:#cbd5e1">AdversarialBrainModeNetwork · K=16 modes · hidden_dim=64 · Upgraded K=32/h=128 training on AMD MI300X</td></tr>
       <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Validation</td><td style="padding:10px 16px;color:#cbd5e1">LOSO AUC = <span style="color:#ef4444;font-weight:600">0.7260</span> · 1,102 held-out subjects · 20 acquisition sites</td></tr>
       <tr style="border-top:1px solid #252a35"><td style="padding:10px 16px;color:#8b95a7;font-size:0.76rem;text-transform:uppercase;letter-spacing:0.5px">Interpretability</td><td style="padding:10px 16px;color:#cbd5e1">Real-time gradient saliency · 7-network aggregation · 3D brain surface</td></tr>
     </table>