Upload folder using huggingface_hub

predict.py

@@ -106,11 +106,11 @@ def load_model(model_dir: str):
     return ens
 
 
-def _predict_8frames_direct(model, context_tensor, last_tensor):
+def _predict_8frames_direct(model, context_tensor, last_tensor, residual_scale=1.0):
     output = model(context_tensor)
     residuals = output.reshape(1, PRED_FRAMES, 3, 64, 64)
     last_expanded = last_tensor.unsqueeze(1).expand_as(residuals)
-    return torch.clamp(last_expanded + residuals, 0, 1)
+    return torch.clamp(last_expanded + residual_scale * residuals, 0, 1)
 
 
 def _predict_ar_frame(model, context_tensor, last_tensor, residual_scale=1.0):
@@ -263,10 +263,10 @@ def predict_next_frame(ens, context_frames: np.ndarray) -> np.ndarray:
             context_tensor = torch.from_numpy(context).to(DEVICE)
             last_tensor = torch.from_numpy(last_frame_t).to(DEVICE)
 
-            predicted_orig = _predict_8frames_direct(ens.models["pole_position"], context_tensor, last_tensor)
+            predicted_orig = _predict_8frames_direct(ens.models["pole_position"], context_tensor, last_tensor, residual_scale=0.97)
             context_flipped = torch.flip(context_tensor, dims=[3])
             last_flipped = torch.flip(last_tensor, dims=[3])
-            predicted_flipped = _predict_8frames_direct(ens.models["pole_position"], context_flipped, last_flipped)
+            predicted_flipped = _predict_8frames_direct(ens.models["pole_position"], context_flipped, last_flipped, residual_scale=0.97)
             predicted_flipped = torch.flip(predicted_flipped, dims=[4])
             predicted = (predicted_orig + predicted_flipped) / 2.0
 

sweep.py

@@ -1,50 +1,50 @@
-"""Sweep rounding bias."""
+"""Sweep PP direct residual scale."""
 import subprocess
 import json
 import re
 
-predict_path = "/home/coder/experiments/2026-04-12-332000-bias-resweep/predict.py"
+predict_path = "/home/coder/experiments/2026-04-12-342000-pp-residual-tune/predict.py"
+
+with open(predict_path, 'r') as f:
+    original = f.read()
 
 results = {}
-for bias in [0.0, 0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.40, 0.50]:
-    with open(predict_path, 'r') as f:
-        content = f.read()
-
-    if bias == 0.0:
-        content = re.sub(
-            r'np\.round\(frame \* 255 \+ [\d.]+\)',
-            'np.round(frame * 255)',
-            content
-        )
-    else:
-        # First handle the case where there's already a bias
-        content = re.sub(
-            r'np\.round\(frame \* 255 \+ [\d.]+\)',
-            f'np.round(frame * 255 + {bias})',
-            content
-        )
-        # Then handle the case where there's no bias (from bias=0.0 step)
-        content = re.sub(
-            r'np\.round\(frame \* 255\)\.clip',
-            f'np.round(frame * 255 + {bias}).clip',
-            content
-        )
+for scale in [0.96, 0.97, 0.98, 0.99, 1.00, 1.01, 1.02, 1.03, 1.04]:
+    content = original
+    # Replace PP direct calls (both orig and flipped)
+    content = re.sub(
+        r'_predict_8frames_direct\(ens\.models\["pole_position"\], context_tensor, last_tensor\)',
+        f'_predict_8frames_direct(ens.models["pole_position"], context_tensor, last_tensor, residual_scale={scale})',
+        content
+    )
+    content = re.sub(
+        r'_predict_8frames_direct\(ens\.models\["pole_position"\], context_flipped, last_flipped\)',
+        f'_predict_8frames_direct(ens.models["pole_position"], context_flipped, last_flipped, 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/2026-04-12-332000-bias-resweep'],
+        ['python', 'task/score.py', '--model_path', '/home/coder/experiments/2026-04-12-342000-pp-residual-tune'],
         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[bias] = data['score']
-            print(f"Bias {bias:.2f}: overall={data['score']:.4f} pong={data['per_game']['pong']['ssim']:.4f} sonic={data['per_game']['sonic']['ssim']:.4f} pp={data['per_game']['pole_position']['ssim']:.4f}")
+            results[scale] = {
+                'score': data['score'],
+                'pp': data['per_game']['pole_position']['ssim']
+            }
+            print(f"Scale {scale}: overall={data['score']:.4f} pp={data['per_game']['pole_position']['ssim']:.4f}")
             break
 
+# Restore
+with open(predict_path, 'w') as f:
+    f.write(original)
+
 print("\n=== Summary ===")
-best_bias = max(results.keys(), key=lambda b: results[b])
-print(f"Best bias: {best_bias} with overall={results[best_bias]:.4f}")
+best_scale = max(results.keys(), key=lambda s: results[s]['pp'])
+print(f"Best PP scale: {best_scale} with pp={results[best_scale]['pp']:.4f}, overall={results[best_scale]['score']:.4f}")