Fix submission.py: per-onnx-file size check, not zip; remove excluded task skip"

neurogolf_solver/submission.py

@@ -8,7 +8,7 @@ import math
 import zipfile
 from collections import Counter
 from .profiler import score_network
-from .constants import MAX_FILESIZE, EXCLUDED_TASKS
+from .constants import MAX_ONNX_FILESIZE, EXCLUDED_TASKS
 
 try:
     import wandb
@@ -30,9 +30,6 @@ def run_tasks(task_nums, tasks, output_dir, providers, conv_budget, excluded_tas
     for tn in task_nums:
         if tn not in tasks:
             continue
-        if tn in excluded_tasks:
-            print(f"Task {tn:3d}: EXCLUDED (officially)")
-            continue
 
         td = tasks[tn]['data']
         ok, sname, sz, t_task, model_path = solve_task(
@@ -47,6 +44,11 @@ def run_tasks(task_nums, tasks, output_dir, providers, conv_budget, excluded_tas
             score = max(1.0, 25.0 - math.log(max(1, cost)))
             total_score += score
 
+            # Check per-file size limit
+            if sz and sz > MAX_ONNX_FILESIZE:
+                print(f"Task {tn:3d}: {sname:25s} OVER SIZE LIMIT ({sz:,} > {MAX_ONNX_FILESIZE:,})")
+                continue
+
             results[tn] = (sname, t_task, sz)
             costs_dict[tn] = cost
             print(f"Task {tn:3d}: {sname:25s} {score:7.3f} {cost:>12} {t_task:7.3f}s  ({sz:>8,} bytes)")
@@ -77,6 +79,14 @@ def generate_submission(output_dir, results, costs_dict, active_tasks):
     total_size = sum(os.path.getsize(os.path.join(output_dir, f))
                      for f in os.listdir(output_dir) if f.endswith('.onnx'))
 
+    # Check per-file size limits
+    oversized = []
+    for f in os.listdir(output_dir):
+        if f.endswith('.onnx'):
+            fsize = os.path.getsize(os.path.join(output_dir, f))
+            if fsize > MAX_ONNX_FILESIZE:
+                oversized.append((f, fsize))
+
     # Create submission.zip
     parent_dir = os.path.dirname(output_dir) or '/kaggle/working/'
     zip_path = os.path.join(parent_dir, 'submission.zip')
@@ -101,16 +111,17 @@ def generate_submission(output_dir, results, costs_dict, active_tasks):
     # Calculate estimated LB score
     unsolved_count = len(active_tasks) - len(results)
     total_score = sum(max(1.0, 25.0 - math.log(max(1, cost))) for cost in costs_dict.values())
-   
+
     return {
         'n_files': n_files,
         'total_size': total_size,
         'zip_path': zip_path,
         'zip_size': zip_size,
         'csv_path': csv_path,
-        'est_lb': total_score,
+        'est_lb': total_score + unsolved_count * 1.0,
         'total_score': total_score,
         'unsolved_count': unsolved_count,
+        'oversized': oversized,
     }
 
 
@@ -119,15 +130,20 @@ def print_summary(results, submission_info, elapsed):
     active_count = submission_info['unsolved_count'] + len(results)
 
     print(f"\n{'=' * 70}")
-    print(f"Solved: {len(results)}/{active_count} active tasks in {elapsed:.0f}s")
+    print(f"Solved: {len(results)}/{active_count} tasks in {elapsed:.0f}s")
     solver_names = [v[0] for v in results.values()]
     sc = Counter(solver_names)
     for s, c in sc.most_common():
         print(f"  {s}: {c}")
 
     print(f"\n{submission_info['n_files']} ONNX files, {submission_info['total_size'] / 1024:.1f} KB uncompressed")
-    zip_ok = 'OK' if submission_info['zip_size'] <= MAX_FILESIZE else 'OVER!'
-    print(f"ZIP size: {submission_info['zip_size'] / 1024:.1f} KB / {MAX_FILESIZE / 1024:.0f} KB limit {zip_ok}")
+    print(f"ZIP size: {submission_info['zip_size'] / 1024:.1f} KB")
+
+    if submission_info['oversized']:
+        print(f"WARNING: {len(submission_info['oversized'])} files exceed 1.44MB limit:")
+        for f, sz in submission_info['oversized']:
+            print(f"  {f}: {sz / 1024:.1f} KB")
+
     print(f"Estimated LB score: {submission_info['est_lb']:.1f} "
-          f"(solved: {submission_info['total_score']:.1f} unsolved: {submission_info['unsolved_count']})")
+          f"(solved: {submission_info['total_score']:.1f} + unsolved: {submission_info['unsolved_count']}x1.0)")
     print(f"Written: {submission_info['zip_path']} | {submission_info['csv_path']}")