Upload ai/utils/check_action_dist.py with huggingface_hub

ai/utils/check_action_dist.py

@@ -0,0 +1,53 @@
+import argparse
+import glob
+
+import numpy as np
+
+
+def analyze_distribution(pattern):
+    files = glob.glob(pattern)
+    if not files:
+        print(f"No files found for {pattern}")
+        return
+
+    total_samples = 0
+    action_counts = {}
+
+    print(f"Analyzing {len(files)} files...")
+    for f in files[:20]:  # Check first 20 chunks for quick insight
+        data = np.load(f)
+        policies = data["policies"]
+
+        # policies is [N, 2000]
+        # Get argmax for each sample
+        best_actions = np.argmax(policies, axis=1)
+
+        total_samples += len(best_actions)
+        unique, counts = np.unique(best_actions, return_counts=True)
+        for a, c in zip(unique, counts):
+            action_counts[a] = action_counts.get(a, 0) + c
+
+    print(f"\nTotal Samples checked: {total_samples}")
+    sorted_actions = sorted(action_counts.items(), key=lambda x: x[1], reverse=True)
+
+    print("\nTop 10 Actions:")
+    for action, count in sorted_actions[:10]:
+        percentage = (count / total_samples) * 100
+        print(f"Action {action:4}: {count:6} samples ({percentage:5.1f}%)")
+
+    # Specifically check Action 0 (End Phase)
+    pass_count = action_counts.get(0, 0)
+    pass_perc = (pass_count / total_samples) * 100
+    print(f"\nAction 0 (Pass/End): {pass_perc:.1f}%")
+
+    # Check "tactical" actions (non-zero)
+    tactical_count = total_samples - pass_count
+    tactical_perc = (tactical_count / total_samples) * 100
+    print(f"Tactical Actions:   {tactical_perc:.1f}%")
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--data", type=str, default="ai/data/self_play_0_chunk_*.npz")
+    args = parser.parse_args()
+    analyze_distribution(args.data)