Upload folder using huggingface_hub

predict.py

@@ -202,12 +202,9 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
             direct_flipped = torch.flip(direct_flipped, dims=[4])
             direct_pred = (direct_orig + direct_flipped) / 2.0
 
-            # Multi-run AR with noise diversity, using direct step 1 as AR init
-            # Get direct model's step 1 prediction for seeding AR context
-            direct_step1 = direct_pred[:, 0]  # [1, 3, 64, 64]
-            direct_step1_flip = torch.flip(direct_step1, dims=[3])
-
+            # Multi-run AR with noise diversity, track TTA disagreement
             all_ar_runs = []
+            tta_disagreements = []  # per-step disagreement from clean run
             for noise_std in [0.0, 1.0/255.0, 2.0/255.0]:
                 ar_preds_run = []
                 ctx = context_tensor.clone()
@@ -220,32 +217,34 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
                     ar_orig = _predict_ar_frame(ens.sonic_ar, ctx_in, last_t)
                     ar_flip = _predict_ar_frame(ens.sonic_ar, ctx_flip_in, last_f)
                     ar_flip_back = torch.flip(ar_flip, dims=[3])
+                    # Track TTA disagreement for clean run
+                    if noise_std == 0:
+                        disagreement = torch.abs(ar_orig - ar_flip_back).mean().item()
+                        tta_disagreements.append(disagreement)
                     ar_frame = (ar_orig + ar_flip_back) / 2.0
                     ar_preds_run.append(ar_frame)
-
-                    # For step 0: use direct model's prediction as context seed
-                    if step == 0:
-                        feed_orig = direct_step1
-                        feed_flip = direct_step1_flip
-                    else:
-                        feed_orig = ar_orig
-                        feed_flip = ar_flip
-
                     ctx_frames = ctx.reshape(1, CONTEXT_FRAMES, 3, 64, 64)
-                    ctx_frames = torch.cat([ctx_frames[:, 1:], feed_orig.unsqueeze(1)], dim=1)
+                    ctx_frames = torch.cat([ctx_frames[:, 1:], ar_orig.unsqueeze(1)], dim=1)
                     ctx = ctx_frames.reshape(1, -1, 64, 64)
-                    last_t = feed_orig
+                    last_t = ar_orig
                     ctx_flip_frames = ctx_flip.reshape(1, CONTEXT_FRAMES, 3, 64, 64)
-                    ctx_flip_frames = torch.cat([ctx_flip_frames[:, 1:], feed_flip.unsqueeze(1)], dim=1)
+                    ctx_flip_frames = torch.cat([ctx_flip_frames[:, 1:], ar_flip.unsqueeze(1)], dim=1)
                     ctx_flip = ctx_flip_frames.reshape(1, -1, 64, 64)
-                    last_f = feed_flip
+                    last_f = ar_flip
                 all_ar_runs.append(torch.stack(ar_preds_run, dim=1))
 
             ar_pred = sum(all_ar_runs) / len(all_ar_runs)
 
             predicted = torch.zeros_like(direct_pred)
             for step in range(PRED_FRAMES):
-                ar_weight = 0.65 - (step / (PRED_FRAMES - 1)) * 0.3
+                base_ar_weight = 0.65 - (step / (PRED_FRAMES - 1)) * 0.3
+                # Adjust based on TTA disagreement
+                disagreement = tta_disagreements[step]
+                if disagreement < 5.0 / 255.0:
+                    base_ar_weight += 0.1  # AR confident
+                elif disagreement > 15.0 / 255.0:
+                    base_ar_weight -= 0.1  # AR uncertain
+                ar_weight = max(0.15, min(0.85, base_ar_weight))
                 direct_weight = 1.0 - ar_weight
                 predicted[:, step] = ar_weight * ar_pred[:, step] + direct_weight * direct_pred[:, step]