| |
| """Summarize metrics from swap_analysis_updated results. |
| |
| Directory layout expected: |
| {root}/short_answer/saved_data/{model_folder}/json/sign_corrected_consistency_{scale}_all_pairs.json |
| {root}/short_answer/saved_data/{model_folder}/csv/delta_similarity_{scale}_L{layer}_all_pairs.csv |
| |
| [Optional – requires separate extraction step] |
| {root}/short_answer/saved_data/{model_folder}/csv/delta_norm_{scale}_L{layer}_all_pairs.csv |
| Expected format: single-column CSV (index = relation label, column = mean norm) |
| ,norm |
| left,12.34 |
| right,11.89 |
| above,9.45 |
| below,9.12 |
| far,7.23 |
| close,7.58 |
| NOTE: delta_similarity CSVs contain cosine similarities only (no magnitude info). |
| To populate Norm columns, save raw delta vector norms during swap_analysis. |
| |
| Each model_folder encodes both the model family and data scale, e.g.: |
| molmo_vanilla, molmo_80k, nvila_800k, nvila_st_800k-st, nvila_synthetic_80k-5pct, qwen_2m |
| |
| Usage: |
| # All models (default) |
| python summarize_metrics_updated.py |
| |
| # Specific models |
| python summarize_metrics_updated.py molmo_2m nvila_800k nvila_st_800k-st |
| """ |
|
|
| import argparse |
| import json |
| import re |
| from pathlib import Path |
|
|
| import numpy as np |
| import pandas as pd |
|
|
| |
| |
| |
|
|
| ROOT = Path("/data/shared/Qwen/experiments/swap_analysis_updated") |
| SAVED_DATA = ROOT / "short_answer" / "saved_data" |
| |
| EXP_DIR = Path("/data/shared/Qwen/experiments") |
|
|
| DISPLAY_NAMES = { |
| "molmo_vanilla": "Molmo vanilla", |
| "molmo_80k": "Molmo 80k", |
| "molmo_400k": "Molmo 400k", |
| "molmo_800k": "Molmo 800k", |
| "molmo_2m": "Molmo 2M", |
| "nvila_vanilla": "NVILA vanilla", |
| "nvila_80k": "NVILA 80k", |
| "nvila_400k": "NVILA 400k", |
| "nvila_800k": "NVILA 800k", |
| "nvila_2m": "NVILA 2M", |
| "nvila_roborefer": "RoboRefer", |
| "nvila_st_80k-st": "NVILA-ST 80k", |
| "nvila_st_400k-st": "NVILA-ST 400k", |
| "nvila_st_800k-st": "NVILA-ST 800k", |
| "nvila_synthetic_80k-5pct": "NVILA Syn 80k-5%", |
| "nvila_synthetic_80k-10pct": "NVILA Syn 80k-10%", |
| "nvila_synthetic_400k-5pct": "NVILA Syn 400k-5%", |
| "nvila_synthetic_800k-5pct": "NVILA Syn 800k-5%", |
| "qwen_vanilla": "Qwen vanilla", |
| "qwen_80k": "Qwen 80k", |
| "qwen_400k": "Qwen 400k", |
| "qwen_800k": "Qwen 800k", |
| "qwen_2m": "Qwen 2M", |
| } |
|
|
| TEXT_FILE_MODEL_NAMES = { |
| "molmo_vanilla": "molmo-7B-O-0924", |
| "molmo_80k": "molmo-7B-O-0924-data_scale_exp_80k", |
| "molmo_400k": "molmo-7B-O-0924-data_scale_exp_400k", |
| "molmo_800k": "molmo-7B-O-0924-data_scale_exp_800k", |
| "molmo_2m": "molmo-7B-O-0924-data_scale_exp_2m", |
| "nvila_vanilla": "NVILA-Lite-2B", |
| "nvila_80k": "NVILA-Lite-2B-data-scale-exp-80k", |
| "nvila_400k": "NVILA-Lite-2B-data-scale-exp-400k", |
| "nvila_800k": "NVILA-Lite-2B-data-scale-exp-800k", |
| "nvila_2m": "NVILA-Lite-2B-data-scale-exp-2m", |
| "nvila_st_80k-st": "NVILA-Lite-2B-ST-80k-5pct", |
| "nvila_st_400k-st": "NVILA-Lite-2B-ST-400k-5pct", |
| "nvila_st_800k-st": "NVILA-Lite-2B-ST-800k-5pct", |
| "nvila_roborefer": "RoboRefer-2B-SFT", |
| "qwen_vanilla": "Qwen2.5-VL-3B-Instruct", |
| "qwen_80k": "Qwen2.5-VL-3B-Instruct-data_scale_exp_80k", |
| "qwen_400k": "Qwen2.5-VL-3B-Instruct-data_scale_exp_400k", |
| "qwen_800k": "Qwen2.5-VL-3B-Instruct-data_scale_exp_800k", |
| "qwen_2m": "Qwen2.5-VL-3B-Instruct-data_scale_exp_2m", |
| } |
|
|
| FOLDER_ORDER = [ |
| "molmo_vanilla", "molmo_80k", "molmo_400k", "molmo_800k", "molmo_2m", |
| "nvila_vanilla", |
| "nvila_80k", "nvila_400k", "nvila_800k", "nvila_2m", |
| "nvila_st_80k-st", "nvila_st_400k-st", "nvila_st_800k-st", |
| "nvila_roborefer", |
| "nvila_synthetic_80k-5pct", "nvila_synthetic_80k-10pct", |
| "nvila_synthetic_400k-5pct", "nvila_synthetic_800k-5pct", |
| "qwen_vanilla", "qwen_80k", "qwen_400k", "qwen_800k", "qwen_2m", |
| ] |
|
|
|
|
| def get_target_layer(folder_name: str) -> int | None: |
| if folder_name.startswith("molmo"): |
| return 23 |
| if folder_name.startswith("nvila"): |
| return 20 |
| if folder_name.startswith("qwen"): |
| return 27 |
| return None |
|
|
|
|
| def make_display_name(folder_name: str) -> str: |
| return DISPLAY_NAMES.get(folder_name, folder_name.replace("_", " ")) |
|
|
|
|
| |
| |
| |
|
|
| def get_peak_consistency(json_file: Path) -> dict: |
| """Peak sign-corrected consistency across all layers per axis.""" |
| with open(json_file) as f: |
| data = json.load(f) |
| result = {} |
| for dim in ("horizontal", "vertical", "distance"): |
| vals = [v["mean"] for k, v in data.items() if k.startswith(f"{dim}_L")] |
| result[dim] = max(vals) if vals else None |
| return result |
|
|
|
|
| def get_layer_consistency(json_file: Path, layer: int) -> dict: |
| """Sign-corrected consistency for horiz/vert/dist at a specific layer. |
| |
| JSON keys follow the pattern: "{dim}_L{layer}" e.g. "distance_L20" |
| Each value is {"mean": float, "std": float, "n": int}. |
| """ |
| with open(json_file) as f: |
| data = json.load(f) |
| result = {} |
| for dim in ("horizontal", "vertical", "distance"): |
| key = f"{dim}_L{layer}" |
| result[dim] = data[key]["mean"] if key in data else None |
| return result |
|
|
|
|
| def _loc(df: pd.DataFrame, row: str, col: str) -> float: |
| """Safe df.loc with below/under alias.""" |
| aliases = {"below": "under", "under": "below"} |
| r = row if row in df.index else aliases.get(row, row) |
| c = col if col in df.columns else aliases.get(col, col) |
| if r not in df.index or c not in df.columns: |
| return float("nan") |
| return float(df.loc[r, c]) |
|
|
|
|
| def get_vd_entanglement(csv_dir: Path, scale: str, layer: int) -> float | None: |
| """VD-entanglement from cosine similarity matrix. |
| |
| Formula: (cos(above,far) + cos(below,close) - cos(above,close) - cos(below,far)) / 4 |
| |
| File: delta_similarity_{scale}_L{layer}_all_pairs.csv |
| Format: 6×6 cosine similarity matrix, labels left/right/above/below/far/close. |
| """ |
| csv_file = csv_dir / f"delta_similarity_{scale}_L{layer}_all_pairs.csv" |
| if not csv_file.exists(): |
| return None |
| df = pd.read_csv(csv_file, index_col=0) |
| vd = ( |
| _loc(df, "above", "far") + _loc(df, "below", "close") |
| - _loc(df, "above", "close") - _loc(df, "below", "far") |
| ) / 4 |
| return float(vd) if np.isfinite(vd) else None |
|
|
|
|
| def get_delta_norms(csv_dir: Path, scale: str, layer: int) -> dict: |
| """Mean delta vector norms for horiz/vert/dist at a specific layer. |
| |
| Requires: delta_norm_{scale}_L{layer}_all_pairs.csv |
| This file is NOT produced by default — it must be generated by saving |
| np.linalg.norm(delta_vec) per relation during swap_analysis. |
| (delta_similarity CSVs contain only cosine similarities, no magnitude info.) |
| |
| Expected format: |
| ,norm |
| left,12.34 right,11.89 above,9.45 below,9.12 far,7.23 close,7.58 |
| |
| Returns all-None dict when file is absent (columns will show N/A). |
| """ |
| norm_file = csv_dir / f"delta_norm_{scale}_L{layer}_all_pairs.csv" |
| if not norm_file.exists(): |
| return {"horizontal": None, "vertical": None, "distance": None} |
|
|
| df = pd.read_csv(norm_file, index_col=0) |
|
|
| def _mean_norm(labels: list[str]) -> float | None: |
| vals = [] |
| for lbl in labels: |
| aliases = {"below": "under", "under": "below"} |
| row = lbl if lbl in df.index else aliases.get(lbl, lbl) |
| if row not in df.index: |
| continue |
| row_vals = df.loc[row].values.astype(float) |
| finite = row_vals[np.isfinite(row_vals)] |
| if finite.size > 0: |
| vals.append(float(np.mean(finite))) |
| return float(np.mean(vals)) if vals else None |
|
|
| return { |
| "horizontal": _mean_norm(["left", "right"]), |
| "vertical": _mean_norm(["above", "below"]), |
| "distance": _mean_norm(["far", "close"]), |
| } |
|
|
|
|
| |
| |
| |
|
|
| def parse_accuracy_text(text_file: Path) -> dict: |
| content = text_file.read_text() |
| sections = re.split(r"={10,}\s*\nModel:\s*", content) |
| result = {} |
| for section in sections[1:]: |
| lines = section.splitlines() |
| model_name = lines[0].strip() |
| consistent = counter = None |
| for line in lines: |
| m = re.match(r"\s*TOTAL\s+consistent\s+(\d+)\s+(\d+)\s+([\d.]+)%", line) |
| if m: |
| consistent = float(m.group(3)) |
| m = re.match(r"\s*TOTAL\s+counter\s+(\d+)\s+(\d+)\s+([\d.]+)%", line) |
| if m: |
| counter = float(m.group(3)) |
| if model_name: |
| result[model_name] = {"consistent": consistent, "counter": counter} |
| return result |
|
|
|
|
| |
| |
| |
|
|
| def fmt(val, fmt_str=".4f", suffix=""): |
| return f"{val:{fmt_str}}{suffix}" if val is not None else "N/A" |
|
|
|
|
| def main(): |
| parser = argparse.ArgumentParser( |
| description="Summarize metrics from swap_analysis_updated results.", |
| formatter_class=argparse.RawDescriptionHelpFormatter, |
| epilog=__doc__, |
| ) |
| parser.add_argument( |
| "models", |
| nargs="*", |
| metavar="MODEL_FOLDER", |
| help="Model folder names under saved_data/ (e.g. molmo_2m nvila_800k). " |
| "Default: all folders found.", |
| ) |
| args = parser.parse_args() |
|
|
| if args.models: |
| model_dirs = [] |
| for name in args.models: |
| d = SAVED_DATA / name |
| if not d.is_dir(): |
| print(f"[warn] Not found, skipping: {d}") |
| else: |
| model_dirs.append(d) |
| else: |
| model_dirs = sorted(d for d in SAVED_DATA.iterdir() if d.is_dir()) |
|
|
| embspatial = parse_accuracy_text(EXP_DIR / "counter_consistent_results_embspatial_all.txt") |
| cvbench3d = parse_accuracy_text(EXP_DIR / "counter_consistent_results_cvbench3d_all.txt") |
|
|
| rows = [] |
| for model_dir in model_dirs: |
| folder_name = model_dir.name |
| json_dir = model_dir / "json" |
| csv_dir = model_dir / "csv" |
|
|
| json_files = list(json_dir.glob("sign_corrected_consistency_*_all_pairs.json")) |
| if not json_files: |
| print(f"[warn] No consistency JSON in {json_dir}, skipping.") |
| continue |
| json_file = json_files[0] |
| m = re.match(r"sign_corrected_consistency_(.+)_all_pairs\.json", json_file.name) |
| if not m: |
| continue |
| scale = m.group(1) |
|
|
| display = make_display_name(folder_name) |
| target_layer = get_target_layer(folder_name) |
|
|
| |
| peak = get_peak_consistency(json_file) |
|
|
| |
| layer_sc = ( |
| get_layer_consistency(json_file, target_layer) |
| if target_layer is not None |
| else {"horizontal": None, "vertical": None, "distance": None} |
| ) |
|
|
| |
| vd_entanglement = ( |
| get_vd_entanglement(csv_dir, scale, target_layer) |
| if (target_layer is not None and csv_dir.is_dir()) |
| else None |
| ) |
|
|
| |
| norms = ( |
| get_delta_norms(csv_dir, scale, target_layer) |
| if (target_layer is not None and csv_dir.is_dir()) |
| else {"horizontal": None, "vertical": None, "distance": None} |
| ) |
|
|
| |
| text_model = TEXT_FILE_MODEL_NAMES.get(folder_name) |
| emb_con = emb_ctr = cvb_con = cvb_ctr = None |
| if text_model: |
| if text_model in embspatial: |
| emb_con = embspatial[text_model]["consistent"] |
| emb_ctr = embspatial[text_model]["counter"] |
| if text_model in cvbench3d: |
| cvb_con = cvbench3d[text_model]["consistent"] |
| cvb_ctr = cvbench3d[text_model]["counter"] |
|
|
| rows.append(dict( |
| folder_name=folder_name, |
| display=display, |
| peak_horiz=peak.get("horizontal"), |
| peak_vert=peak.get("vertical"), |
| peak_dist=peak.get("distance"), |
| target_layer=target_layer, |
| layer_horiz=layer_sc["horizontal"], |
| layer_vert=layer_sc["vertical"], |
| layer_dist=layer_sc["distance"], |
| vd_entanglement=vd_entanglement, |
| norm_horiz=norms["horizontal"], |
| norm_vert=norms["vertical"], |
| norm_dist=norms["distance"], |
| emb_con=emb_con, emb_ctr=emb_ctr, |
| cvb_con=cvb_con, cvb_ctr=cvb_ctr, |
| )) |
|
|
| if not rows: |
| print("No data found.") |
| return |
|
|
| rows.sort(key=lambda r: FOLDER_ORDER.index(r["folder_name"]) |
| if r["folder_name"] in FOLDER_ORDER else 99) |
|
|
| records = [] |
| for r in rows: |
| layer = r["target_layer"] if r["target_layer"] is not None else "?" |
| records.append({ |
| "Model": r["display"], |
| "Peak Horiz": fmt(r["peak_horiz"]), |
| "Peak Vert": fmt(r["peak_vert"]), |
| "Peak Dist": fmt(r["peak_dist"]), |
| "Ent. Layer": str(layer), |
| "Layer Horiz SC": fmt(r["layer_horiz"]), |
| "Layer Vert SC": fmt(r["layer_vert"]), |
| "Layer Dist SC": fmt(r["layer_dist"]), |
| "VD-Entanglement": fmt(r["vd_entanglement"]), |
| "Norm Horiz": fmt(r["norm_horiz"]), |
| "Norm Vert": fmt(r["norm_vert"]), |
| "Norm Dist": fmt(r["norm_dist"]), |
| "EmbSpatial (con)": fmt(r["emb_con"], ".1f", "%"), |
| "EmbSpatial (ctr)": fmt(r["emb_ctr"], ".1f", "%"), |
| "CVBench3D (con)": fmt(r["cvb_con"], ".1f", "%"), |
| "CVBench3D (ctr)": fmt(r["cvb_ctr"], ".1f", "%"), |
| }) |
|
|
| df = pd.DataFrame(records) |
| print(df.to_string(index=False)) |
|
|
| csv_path = EXP_DIR / "summarize_metrics" / "swap_analysis_updated" / "short_answer_including_norm.csv" |
| csv_path.parent.mkdir(parents=True, exist_ok=True) |
| df.to_csv(csv_path, index=False) |
| print(f"\nSaved: {csv_path}") |
|
|
|
|
| if __name__ == "__main__": |
| main() |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
| |
|
|
| |
| |
| |
|
|
| |
| |
|
|
| |
| |
| |
|
|
| |
| |
| |
|
|
| |
| |
| |
| |
|
|
| |
| |
|
|
| |
| |
| |
|
|
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
|
|
| |
| |
| |
|
|
|
|
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
|
|
|
|
| |
| |
| |
| |
| |
| |
| |
|
|
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
|
|
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
|
|
| |
| |
| |
|
|
| |
| |
|
|
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| |
| |
| |
|
|
| |
| |
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| |
| |
|
|
| |
| |
|
|
| |
| |
| |
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| |
| |
| |
|
|
| |
| |
|
|
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
| |
|
|
| |
| |
|
|
| |
| |
| |
| |
| |
|
|
|
|
| |
| |
|
|
