| |
| |
|
|
| import argparse |
| import json |
| import os |
| from glob import glob |
|
|
|
|
| def load_json(path): |
| with open(path, "r") as handle: |
| return json.load(handle) |
|
|
|
|
| def summarize_trace(path): |
| payload = load_json(path) |
| metrics = payload["metrics"] |
| base = metrics["base"] |
| best_name, best_metrics = max( |
| [(name, values) for name, values in metrics.items() if name != "base"], |
| key=lambda item: item[1]["mAP"], |
| ) |
| oracle = payload.get("oracle", {}) |
| usage = oracle.get("best_candidate_usage", {}) |
| base_best = usage.get("base", 0) |
| query_count = oracle.get("query_count", 0) |
| return { |
| "dataset": payload.get("dataset", os.path.basename(path)), |
| "base_rank1": base["Rank-1"], |
| "base_map": base["mAP"], |
| "best_fixed": best_name, |
| "best_fixed_map": best_metrics["mAP"], |
| "oracle_rank1": oracle.get("oracle_Rank1_by_query_rank1", 0.0), |
| "oracle_map": oracle.get("oracle_mAP_by_query_ap", 0.0), |
| "improved_queries": oracle.get("queries_improved_over_base", 0), |
| "query_count": query_count, |
| "base_best": base_best, |
| } |
|
|
|
|
| def summarize_exchange(path, topk=8): |
| payload = load_json(path) |
| rows = [] |
| for name, values in payload.get("results", {}).items(): |
| row = {"candidate": name} |
| row.update(values) |
| rows.append(row) |
| rows.sort(key=lambda item: item["ownership_gap"], reverse=True) |
| return { |
| "path": path, |
| "num_samples": payload.get("num_samples", 0), |
| "valid_anchors": payload.get("valid_anchors", 0), |
| "top_positive": rows[:topk], |
| "top_negative": list(reversed(rows[-topk:])), |
| } |
|
|
|
|
| def summarize_crfa(path): |
| payload = load_json(path) |
| history = payload.get("history", []) |
| if not history: |
| return None |
| last = history[-1] |
| loss_cfg = payload.get("loss_cfg", {}) |
| weights = loss_cfg.get("weights") or payload.get("loss_weights", {}) |
| mode = loss_cfg.get("mode") |
| if mode is None and weights: |
| mode = "softplus" |
| return { |
| "file": os.path.basename(path), |
| "mode": mode or "unknown", |
| "same_weight": weights.get("same", 0.0), |
| "diff_weight": weights.get("diff", 0.0), |
| "view_weight": weights.get("view", 0.0), |
| "iter": last.get("iter", payload.get("max_iters", 0)), |
| "loss": last.get("loss", 0.0), |
| "ownership_gap": last.get("ownership_gap", 0.0), |
| "same_gain": last.get("same_gain", 0.0), |
| "diff_gain": last.get("diff_gain", 0.0), |
| "view_shift_abs": last.get("view_shift_abs", 0.0), |
| } |
|
|
|
|
| def summarize_crfa_eval(path): |
| payload = load_json(path) |
| rows = [] |
| for name, values in payload.get("summary", {}).items(): |
| row = {"file": os.path.basename(path), "candidate": name} |
| row.update(values) |
| rows.append(row) |
| return rows |
|
|
|
|
| def print_trace_summary(rows): |
| print("Trace diagnostics") |
| print( |
| "dataset,base_rank1,base_mAP,best_fixed,best_fixed_mAP," |
| "oracle_rank1,oracle_mAP,improved_queries,base_best" |
| ) |
| for row in rows: |
| print( |
| f"{row['dataset']},{row['base_rank1']:.2f},{row['base_map']:.2f}," |
| f"{row['best_fixed']},{row['best_fixed_map']:.2f}," |
| f"{row['oracle_rank1']:.2f},{row['oracle_map']:.2f}," |
| f"{row['improved_queries']}/{row['query_count']}," |
| f"{row['base_best']}/{row['query_count']}" |
| ) |
|
|
|
|
| def print_exchange_summary(summary): |
| print("\nExchange diagnostics") |
| print(f"path={summary['path']}") |
| print(f"samples={summary['num_samples']} valid_anchors={summary['valid_anchors']}") |
| print("top positive ownership gaps:") |
| for row in summary["top_positive"]: |
| print( |
| f" {row['candidate']}: gap={row['ownership_gap']:.6f}, " |
| f"same={row['delta_same_mean']:.6f}, diff={row['delta_diff_mean']:.6f}, " |
| f"gap_pos={row['gap_positive_ratio']:.3f}" |
| ) |
| print("top negative ownership gaps:") |
| for row in summary["top_negative"]: |
| print( |
| f" {row['candidate']}: gap={row['ownership_gap']:.6f}, " |
| f"same={row['delta_same_mean']:.6f}, diff={row['delta_diff_mean']:.6f}, " |
| f"gap_pos={row['gap_positive_ratio']:.3f}" |
| ) |
|
|
|
|
| def print_crfa_summary(rows): |
| print("\nCRFA training diagnostics") |
| print("file,mode,same_w,diff_w,view_w,iter,loss,ownership_gap,same_gain,diff_gain,view_shift_abs") |
| for row in rows: |
| print( |
| f"{row['file']},{row['mode']},{row['same_weight']:.2f},{row['diff_weight']:.2f}," |
| f"{row['view_weight']:.2f},{row['iter']:.0f},{row['loss']:.6f}," |
| f"{row['ownership_gap']:.6f},{row['same_gain']:.6f}," |
| f"{row['diff_gain']:.6f},{row['view_shift_abs']:.6f}" |
| ) |
|
|
|
|
| def print_crfa_eval_summary(rows): |
| print("\nCRFA held-out exchange diagnostics") |
| print("file,candidate,valid_anchors,ownership_gap,same_gain,diff_gain,gap_pos,same_pos,diff_nonpos") |
| for row in rows: |
| print( |
| f"{row['file']},{row['candidate']},{row['valid_anchors']}," |
| f"{row['ownership_gap']:.6f},{row['same_gain']:.6f}," |
| f"{row['diff_gain']:.6f},{row['gap_positive_ratio']:.3f}," |
| f"{row['same_positive_ratio']:.3f},{row['diff_nonpositive_ratio']:.3f}" |
| ) |
|
|
|
|
| def parse_args(): |
| parser = argparse.ArgumentParser(description="Summarize CAVI diagnostic JSON files") |
| parser.add_argument("--dir", default="logs/CARGO/CAVI_TRACE") |
| parser.add_argument("--exchange", default=None) |
| parser.add_argument("--crfa-glob", default="crfa_*.json") |
| parser.add_argument("--crfa-eval-glob", default="crfa_eval*.json") |
| parser.add_argument("--topk", type=int, default=8) |
| return parser.parse_args() |
|
|
|
|
| def main(): |
| args = parse_args() |
| trace_paths = sorted( |
| path for path in glob(os.path.join(args.dir, "*_trace.json")) |
| if not os.path.basename(path).startswith("smoke") |
| ) |
| if trace_paths: |
| print_trace_summary([summarize_trace(path) for path in trace_paths]) |
| exchange_path = args.exchange or os.path.join(args.dir, "exchange_smoke.json") |
| if os.path.exists(exchange_path): |
| print_exchange_summary(summarize_exchange(exchange_path, args.topk)) |
| crfa_rows = [] |
| for path in sorted(glob(os.path.join(args.dir, args.crfa_glob))): |
| row = summarize_crfa(path) |
| if row is not None: |
| crfa_rows.append(row) |
| if crfa_rows: |
| print_crfa_summary(crfa_rows) |
| eval_rows = [] |
| for path in sorted(glob(os.path.join(args.dir, args.crfa_eval_glob))): |
| eval_rows.extend(summarize_crfa_eval(path)) |
| if eval_rows: |
| print_crfa_eval_summary(eval_rows) |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
