Fix: mLSTM SiLU gate+activation, GroupNorm 192, stochastic depth 0.05, Hanning window

test_all.py

@@ -73,6 +73,10 @@ def test_mlstm_cell():
     assert cell_params < 1_000_000, f"Cell has {cell_params:,} params (should be <1M)"
     assert cell_params > 800_000, f"Cell has {cell_params:,} params (should be >800K)"
     
+    # Verify GroupNorm uses 192 groups (num_proj_heads), not 4 (num_heads)
+    assert cell.outnorm.num_groups == 192, f"GroupNorm should have 192 groups, got {cell.outnorm.num_groups}"
+    print(f"  GroupNorm groups: {cell.outnorm.num_groups} (correct: per-projection-head)")
+    
     x = torch.randn(2, 20, 384)
     y = cell(x)
     assert y.shape == (2, 20, 384), f"Cell output shape: {y.shape}"
@@ -190,7 +194,7 @@ test("Backbone (TMoE + integrated FiLM)", test_backbone_tmoe_film)
 # Test 6: Prediction Heads
 # ============================================================
 def test_heads():
-    from vil_tracker.models.heads import CenterHead, UncertaintyHead, decode_predictions
+    from vil_tracker.models.heads import CenterHead, UncertaintyHead, decode_predictions, create_hanning_window
     
     center_head = CenterHead(dim=384, feat_size=16)
     unc_head = UncertaintyHead(dim=384, feat_size=16)
@@ -205,11 +209,24 @@ def test_heads():
     assert preds['size'].shape == (2, 2, 16, 16), f"Size shape: {preds['size'].shape}"
     assert preds['offset'].shape == (2, 2, 16, 16), f"Offset shape: {preds['offset'].shape}"
     
-    # Decode
+    # Decode without Hanning
     boxes, scores = decode_predictions(preds['heatmap'], preds['size'], preds['offset'])
     assert boxes.shape == (2, 4), f"Boxes shape: {boxes.shape}"
     assert scores.shape == (2,), f"Scores shape: {scores.shape}"
     
+    # Decode WITH Hanning window
+    hann = create_hanning_window(16)
+    assert hann.shape == (16, 16), f"Hanning shape: {hann.shape}"
+    assert abs(hann[8, 8].item() - 1.0) < 0.05, f"Hanning center should be ~1.0, got {hann[8, 8]}"
+    assert hann[0, 0].item() < 0.01, f"Hanning corner should be ~0, got {hann[0, 0]}"
+    
+    boxes_h, scores_h = decode_predictions(preds['heatmap'], preds['size'], preds['offset'],
+                                           hanning_window=hann)
+    assert boxes_h.shape == (2, 4), f"Hanning boxes shape: {boxes_h.shape}"
+    print(f"  Hanning window: center={hann[8,8]:.3f}, corner={hann[0,0]:.6f}")
+    print(f"  Without Hanning: box={boxes[0].tolist()}, score={scores[0].item():.4f}")
+    print(f"  With Hanning:    box={boxes_h[0].tolist()}, score={scores_h[0].item():.4f}")
+    
     # Uncertainty
     log_var = unc_head(search_feat)
     assert log_var.shape == (2, 1, 16, 16), f"Log variance shape: {log_var.shape}"