Upload folder using huggingface_hub

predict.py

@@ -202,43 +202,58 @@ 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
+            # Shift helpers using roll + edge replication
+            def shift_right(t):
+                shifted = torch.roll(t, 1, dims=-1)
+                shifted[:, :, :, 0] = shifted[:, :, :, 1]
+                return shifted
+            def shift_left(t):
+                shifted = torch.roll(t, -1, dims=-1)
+                shifted[:, :, :, -1] = shifted[:, :, :, -2]
+                return shifted
+            def unshift_right(t):
+                return shift_left(t)
+            def unshift_left(t):
+                return shift_right(t)
+
+            # Build augmentation list: (context_aug, last_aug, undo_fn)
+            ctx_sr = shift_right(context_tensor)
+            last_sr = shift_right(last_tensor)
+            ctx_sl = shift_left(context_tensor)
+            last_sl = shift_left(last_tensor)
+
+            augmentations = [
+                (context_tensor, last_tensor, lambda x: x),
+                (context_flipped, last_flipped, lambda x: torch.flip(x, dims=[3])),
+                (ctx_sr, last_sr, unshift_right),
+                (ctx_sl, last_sl, unshift_left),
+            ]
+
+            # Multi-run AR with noise diversity x augmentations
             all_ar_runs = []
             for noise_std in [0.0, 1.0/255.0, 2.0/255.0]:
-                ar_preds_run = []
-                ctx = context_tensor.clone()
-                ctx_flip = context_flipped.clone()
-                last_t = last_tensor.clone()
-                last_f = last_flipped.clone()
-                for step in range(PRED_FRAMES):
-                    ctx_in = ctx if noise_std == 0 else torch.clamp(ctx + torch.randn_like(ctx) * noise_std, 0, 1)
-                    ctx_flip_in = ctx_flip if noise_std == 0 else torch.clamp(ctx_flip + torch.randn_like(ctx_flip) * noise_std, 0, 1)
-                    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])
-                    ar_frame = (ar_orig + ar_flip_back) / 2.0
-                    ar_preds_run.append(ar_frame)
-                    ctx_frames = ctx.reshape(1, CONTEXT_FRAMES, 3, 64, 64)
-                    ctx_frames = torch.cat([ctx_frames[:, 1:], ar_orig.unsqueeze(1)], dim=1)
-                    ctx = ctx_frames.reshape(1, -1, 64, 64)
-                    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:], ar_flip.unsqueeze(1)], dim=1)
-                    ctx_flip = ctx_flip_frames.reshape(1, -1, 64, 64)
-                    last_f = ar_flip
-                all_ar_runs.append(torch.stack(ar_preds_run, dim=1))
+                for ctx_aug, last_aug, undo_fn in augmentations:
+                    ar_preds_run = []
+                    ctx = ctx_aug.clone()
+                    last_t = last_aug.clone()
+                    for step in range(PRED_FRAMES):
+                        ctx_in = ctx if noise_std == 0 else torch.clamp(ctx + torch.randn_like(ctx) * noise_std, 0, 1)
+                        ar_out = _predict_ar_frame(ens.sonic_ar, ctx_in, last_t)
+                        ar_frame = undo_fn(ar_out)
+                        ar_preds_run.append(ar_frame)
+                        ctx_frames = ctx.reshape(1, CONTEXT_FRAMES, 3, 64, 64)
+                        ctx_frames = torch.cat([ctx_frames[:, 1:], ar_out.unsqueeze(1)], dim=1)
+                        ctx = ctx_frames.reshape(1, -1, 64, 64)
+                        last_t = ar_out
+                    all_ar_runs.append(torch.stack(ar_preds_run, dim=1))
 
             ar_pred = sum(all_ar_runs) / len(all_ar_runs)
 
-            # Agreement-based blending: trust AR where models agree
             predicted = torch.zeros_like(direct_pred)
-            low_thresh = 10.0 / 255.0
-            high_thresh = 50.0 / 255.0
             for step in range(PRED_FRAMES):
-                disagreement = torch.abs(ar_pred[:, step] - direct_pred[:, step])
-                # Where agree: AR weight=1.0, where disagree strongly: AR weight=0.5
-                ar_w = 1.0 - 0.5 * torch.clamp((disagreement - low_thresh) / (high_thresh - low_thresh), 0, 1)
-                predicted[:, step] = ar_w * ar_pred[:, step] + (1.0 - ar_w) * direct_pred[:, step]
+                ar_weight = 0.65 - (step / (PRED_FRAMES - 1)) * 0.3
+                direct_weight = 1.0 - ar_weight
+                predicted[:, step] = ar_weight * ar_pred[:, step] + direct_weight * direct_pred[:, step]
 
         predicted_np = predicted[0].cpu().numpy()
         ens.direct_cache = []