Upload folder using huggingface_hub

predict.py

@@ -155,7 +155,7 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
             ctx = context_tensor.clone()
             last_t = last_tensor.clone()
             for step in range(PRED_FRAMES):
-                predicted = _predict_ar_frame(ens.models["pong"], ctx, last_t, residual_scale=1.03)
+                predicted = _predict_ar_frame(ens.models["pong"], ctx, last_t, residual_scale=1.02)
                 ar_preds.append(predicted)
                 ctx_frames = ctx.reshape(1, CONTEXT_FRAMES, 3, 64, 64)
                 ctx_frames = torch.cat([ctx_frames[:, 1:], predicted.unsqueeze(1)], dim=1)
@@ -174,7 +174,7 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
         ens.direct_cache = []
         for i in range(PRED_FRAMES):
             frame = np.transpose(predicted_np[i], (1, 2, 0))
-            frame = np.round(frame * 255 + 0.3).clip(0, 255).astype(np.uint8)
+            frame = np.round(frame * 255 + 0.1).clip(0, 255).astype(np.uint8)
             ens.direct_cache.append(frame)
 
         result = ens.direct_cache[ens.cache_step]
@@ -202,8 +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
+            # Multi-run AR with noise diversity (fixed seed for reproducibility)
             all_ar_runs = []
+            torch.manual_seed(2)
             for noise_std in [0.0, 1.0/255.0, 2.0/255.0]:
                 ar_preds_run = []
                 ctx = context_tensor.clone()
@@ -240,7 +241,7 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
         ens.direct_cache = []
         for i in range(PRED_FRAMES):
             frame = np.transpose(predicted_np[i], (1, 2, 0))
-            frame = np.round(frame * 255 + 0.2).clip(0, 255).astype(np.uint8)
+            frame = np.round(frame * 255 + 0.1).clip(0, 255).astype(np.uint8)
             ens.direct_cache.append(frame)
 
         result = ens.direct_cache[ens.cache_step]
@@ -272,7 +273,7 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
         ens.direct_cache = []
         for i in range(PRED_FRAMES):
             frame = np.transpose(predicted_np[i], (1, 2, 0))
-            frame = np.round(frame * 255).clip(0, 255).astype(np.uint8)
+            frame = np.round(frame * 255 + 0.1).clip(0, 255).astype(np.uint8)
             ens.direct_cache.append(frame)
 
         result = ens.direct_cache[ens.cache_step]

sweep.py

@@ -0,0 +1,42 @@
+"""Sweep Pong AR residual scale."""
+import subprocess
+import json
+import re
+
+predict_path = "/home/coder/experiments/2026-04-12-330000-pong-amp-sweep/predict.py"
+
+results = {}
+for scale in [1.01, 1.02, 1.03, 1.04, 1.05, 1.06, 1.07]:
+    with open(predict_path, 'r') as f:
+        content = f.read()
+
+    # Replace Pong AR residual_scale (only in Pong section, identified by "pong" model)
+    content = re.sub(
+        r'_predict_ar_frame\(ens\.models\["pong"\], ctx, last_t, residual_scale=[\d.]+\)',
+        f'_predict_ar_frame(ens.models["pong"], ctx, last_t, residual_scale={scale})',
+        content
+    )
+
+    with open(predict_path, 'w') as f:
+        f.write(content)
+
+    result = subprocess.run(
+        ['python', 'task/score.py', '--model_path', '/home/coder/experiments/2026-04-12-330000-pong-amp-sweep'],
+        capture_output=True, text=True, cwd='/home/coder'
+    )
+
+    for line in result.stdout.strip().split('\n'):
+        if '"score"' in line:
+            data = json.loads(line)
+            results[scale] = {
+                'score': data['score'],
+                'pong': data['per_game']['pong']['ssim'],
+                'sonic': data['per_game']['sonic']['ssim'],
+                'pp': data['per_game']['pole_position']['ssim']
+            }
+            print(f"Scale {scale}: overall={data['score']:.4f} pong={data['per_game']['pong']['ssim']:.4f}")
+            break
+
+print("\n=== Summary ===")
+best_scale = max(results.keys(), key=lambda s: results[s]['pong'])
+print(f"Best Pong scale: {best_scale} with pong={results[best_scale]['pong']:.4f}, overall={results[best_scale]['score']:.4f}")