analyze_counter_consistent.py

#!/usr/bin/env python3
"""
Counter vs Consistent Example Analysis Script

2D Heuristic (shared across datasets):
  Upper part of image (small y) = farther from camera
  Lower part of image (large y) = closer to camera

Datasets:
  embspatial (default):
    FAR/CLOSE questions in EmbSpatial-Bench
    Consistent: GT answer agrees with the 2D heuristic (Height-Depth Entanglement)
    Counter: GT answer contradicts the 2D heuristic

  cvbench3d:
    Depth questions: "Which object is closer to the camera?"
      Consistent: GT object (closer) has larger center_y (lower in image)
      Counter: GT object (closer) has smaller center_y (higher in image)
    Distance questions: "Which object is closer to [reference]?"
      2D heuristic: smaller pixel distance to reference = closer in 3D
      Consistent: GT candidate has smaller 2D pixel distance to reference
      Counter: GT candidate has larger 2D pixel distance to reference

Usage:
    python experiments/analyze_counter_consistent.py <model_result.xlsx> [--verbose]
    python experiments/analyze_counter_consistent.py --compare <file1.xlsx> <file2.xlsx> ...
    python experiments/analyze_counter_consistent.py --dataset cvbench3d <result.xlsx>
    python experiments/analyze_counter_consistent.py --dataset cvbench3d --compare <file1.xlsx> ...
"""

import argparse
import ast
import pandas as pd
import numpy as np
from datasets import load_dataset
from pathlib import Path
from typing import Dict, List, Tuple, Optional
import json
import sys


class TeeWriter:
    """Write stdout to both terminal and file simultaneously"""
    def __init__(self, filepath):
        self.terminal = sys.stdout
        self.file = open(filepath, 'w', encoding='utf-8')

    def write(self, message):
        self.terminal.write(message)
        self.file.write(message)
        self.file.flush()

    def flush(self):
        self.terminal.flush()
        self.file.flush()

    def close(self):
        self.file.close()
        return self.terminal


# =============================================================================
# EmbSpatial-Bench
# =============================================================================

def get_bbox_center_y(bbox: List[int], source: str = None) -> float:
    """
    BBox -> center y coordinate, format varies by source:
      ScanNet / MP3D : [x1, y1, w,  h ] -> y1 + h/2
      AI2Thor        : [x1, y1, x2, y2] -> (y1 + y2) / 2
    """
    if source == 'ai2thor':
        return (bbox[1] + bbox[3]) / 2
    else:
        return bbox[1] + bbox[3] / 2


def classify_sample(relation: str, objects: Dict, gt_answer_idx: int,
                    answer_options: List[str] = None,
                    image_height: int = None, threshold_ratio: float = 0.05,
                    data_source: str = None) -> Tuple[str, Dict]:
    """
    Classify a sample as Consistent / Counter / Ambiguous.

    Args:
        relation: 'far' or 'close'
        objects: {'bbox': [...], 'name': [...]}
        gt_answer_idx: GT answer index (0-based, relative to answer_options)
        answer_options: list of answer choices (used to match bbox by name)
        image_height: image height for threshold normalization (pass PIL image.size[1])
        threshold_ratio: ambiguous decision threshold as a fraction of image height
        data_source: 'scannet' | 'mp3d' | 'ai2thor' (selects bbox format)

    Returns:
        classification: 'consistent', 'counter', or 'ambiguous'
        details: dict with classification details
    """
    if relation not in ['far', 'close']:
        return 'not_applicable', {}

    bboxes = objects['bbox']
    names = objects['name']

    if len(bboxes) < 2:
        return 'insufficient_objects', {}

    # answer_options and objects['name'] may differ (e.g. 'Unknown')
    # resolve GT answer index against objects['name']
    if answer_options is not None and gt_answer_idx < len(answer_options):
        gt_answer_name = answer_options[gt_answer_idx]
        if gt_answer_name in names:
            gt_answer_idx = names.index(gt_answer_name)
        elif gt_answer_name == 'Unknown' or gt_answer_idx >= len(bboxes):
            return 'unknown_object', {}

    # bounds check
    if gt_answer_idx >= len(bboxes):
        return 'index_out_of_range', {}

    # compute center y per object using source-specific bbox format
    center_ys = [get_bbox_center_y(bbox, source=data_source) for bbox in bboxes]

    # GT object center y
    gt_center_y = center_ys[gt_answer_idx]

    # mean center y of all other objects
    other_ys = [y for i, y in enumerate(center_ys) if i != gt_answer_idx]
    other_avg_y = np.mean(other_ys)

    # y difference
    y_diff = gt_center_y - other_avg_y

    # threshold normalized by image height
    if image_height:
        threshold = image_height * threshold_ratio
    else:
        threshold = 20  # fallback: 20 pixels

    details = {
        'gt_object': names[gt_answer_idx],
        'gt_center_y': gt_center_y,
        'other_avg_y': other_avg_y,
        'y_diff': y_diff,
        'threshold': threshold,
        'all_objects': list(zip(names, center_ys))
    }

    # ambiguous if difference is too small
    if abs(y_diff) < threshold:
        return 'ambiguous', details

    # FAR: consistent if GT is higher (smaller y)
    if relation == 'far':
        if gt_center_y < other_avg_y:
            return 'consistent', details
        else:
            return 'counter', details

    # CLOSE: consistent if GT is lower (larger y)
    else:
        if gt_center_y > other_avg_y:
            return 'consistent', details
        else:
            return 'counter', details


def get_image_height_by_source(data_source: str) -> int:
    """Return fallback image height by data source (used when PIL image is unavailable)"""
    heights = {
        'ai2thor': 300,
        'mp3d': 480,
        'scannet': 968,
    }
    return heights.get(data_source, 480)


def build_classification_cache(verbose: bool = False) -> Dict[str, Dict]:
    """
    Build a counter/consistent classification cache for the full EmbSpatial-Bench dataset.
    """
    print("Loading EmbSpatial-Bench dataset...")
    ds = load_dataset('FlagEval/EmbSpatial-Bench', split='test')

    cache = {}
    stats = {'far': {'consistent': 0, 'counter': 0, 'ambiguous': 0},
             'close': {'consistent': 0, 'counter': 0, 'ambiguous': 0}}

    for item in ds:
        question_id = item['question_id']
        relation = item['relation']

        if relation not in ['far', 'close']:
            cache[question_id] = {'classification': 'not_applicable', 'relation': relation}
            continue

        objects = item['objects']
        gt_answer_idx = item['answer']  # 0-based index
        answer_options = item['answer_options']
        data_source = item['data_source']

        # use actual image height from PIL image (image.size -> (width, height))
        pil_image = item.get('image')
        if pil_image is not None and hasattr(pil_image, 'size'):
            image_height = pil_image.size[1]
        else:
            image_height = get_image_height_by_source(data_source)

        classification, details = classify_sample(
            relation, objects, gt_answer_idx, answer_options, image_height,
            data_source=data_source
        )

        cache[question_id] = {
            'classification': classification,
            'relation': relation,
            'data_source': item['data_source'],
            'details': details
        }

        if relation in stats and classification in stats[relation]:
            stats[relation][classification] += 1

    if verbose:
        print("\n=== Classification Statistics ===")
        for rel in ['far', 'close']:
            total = sum(stats[rel].values())
            print(f"\n{rel.upper()} (n={total}):")
            for cls, cnt in stats[rel].items():
                pct = cnt / total * 100 if total > 0 else 0
                print(f"  {cls}: {cnt} ({pct:.1f}%)")

    return cache


def analyze_embspatial_results(xlsx_path: str, cache: Dict[str, Dict],
                               verbose: bool = False) -> Tuple[Dict, List[Dict]]:
    """Analyze a model result xlsx file against the EmbSpatialBench classification cache."""
    df = pd.read_excel(xlsx_path)

    results = {
        'far': {
            'consistent': {'correct': 0, 'total': 0},
            'counter': {'correct': 0, 'total': 0},
            'ambiguous': {'correct': 0, 'total': 0}
        },
        'close': {
            'consistent': {'correct': 0, 'total': 0},
            'counter': {'correct': 0, 'total': 0},
            'ambiguous': {'correct': 0, 'total': 0}
        }
    }

    counter_examples = []

    for _, row in df.iterrows():
        question_id = row['question_id']
        category = row['category']
        hit = row['hit']

        if category not in ['far', 'close']:
            continue

        if question_id not in cache:
            continue

        info = cache[question_id]
        classification = info['classification']

        if classification not in ['consistent', 'counter', 'ambiguous']:
            continue

        results[category][classification]['total'] += 1
        if hit == 1:
            results[category][classification]['correct'] += 1

        if classification == 'counter':
            counter_examples.append({
                'question_id': question_id,
                'relation': category,
                'hit': hit,
                'prediction': row['prediction'],
                'answer': row['answer'],
                'data_source': info['data_source'],
                'details': info.get('details', {})
            })

    return results, counter_examples


# =============================================================================
# CV-Bench-3D
# =============================================================================

# Known image heights per source dataset (used for threshold normalization)
# Omni3D_SUNRGBD has variable sizes; fallback to max bbox y2 estimate.
_CVBENCH3D_SOURCE_HEIGHTS = {
    'Omni3D_Hypersim': 768,
    'Omni3D_nuScenes': 900,
}


def classify_cvbench3d_row(row, depth_threshold_ratio: float = 0.05) -> Tuple[str, Dict]:
    """
    Classify a single CV-Bench-3D row as consistent / counter / ambiguous.

    Only Depth questions are classified — they share the same height-depth
    entanglement heuristic as EmbSpatial-Bench:
        2D heuristic: lower in image (larger center_y) = closer to camera
        Consistent: GT object (closer to camera) has larger center_y
        Counter:     GT object (closer to camera) has smaller center_y

    Distance questions ask "which object is closer to [reference] in 3D real-world
    distance?" — this is inter-object 3D distance, not viewer distance.  No
    equivalent 2D projection heuristic exists (height-depth entanglement does not
    apply), so Distance rows are always marked 'not_applicable'.
    """
    category = row['category']
    answer_letter = str(row['answer']).strip()

    if category != 'Depth':
        return 'not_applicable', {}

    try:
        bbox_list = ast.literal_eval(row['bbox'])
    except (ValueError, SyntaxError):
        return 'invalid_bbox', {}

    if len(bbox_list) != 2:
        return 'invalid_bbox', {}

    cy_A = (bbox_list[0][1] + bbox_list[0][3]) / 2
    cy_B = (bbox_list[1][1] + bbox_list[1][3]) / 2

    gt_y = cy_A if answer_letter == 'A' else cy_B
    other_y = cy_B if answer_letter == 'A' else cy_A
    y_diff = gt_y - other_y  # positive = GT is lower in image

    # Estimate image height: prefer known source height, fall back to max bbox y2
    source_dataset = str(row.get('source_dataset', ''))
    known_h = _CVBENCH3D_SOURCE_HEIGHTS.get(source_dataset, 0)
    est_h = max(bb[3] for bb in bbox_list)
    image_height = max(known_h, est_h)
    threshold = image_height * depth_threshold_ratio

    details = {
        'answer': answer_letter,
        'center_y_A': cy_A,
        'center_y_B': cy_B,
        'y_diff': y_diff,
        'threshold': threshold,
        'image_height_est': image_height,
        'source_dataset': source_dataset,
    }

    if abs(y_diff) < threshold:
        return 'ambiguous', details
    # Consistent: GT (closer to camera) is lower in image (larger y)
    return ('consistent' if gt_y > other_y else 'counter'), details


def analyze_cvbench3d_results(xlsx_path: str, verbose: bool = False,
                              depth_threshold_ratio: float = 0.05) -> Tuple[Dict, List[Dict]]:
    """
    Analyze a CV-Bench-3D result xlsx file.

    Only the Depth category is classified into consistent / counter / ambiguous,
    because it shares the height-depth entanglement heuristic with EmbSpatial-Bench.
    Distance (inter-object 3D distance) has no analogous 2D projection heuristic
    and is excluded from the consistent/counter analysis.
    """
    df = pd.read_excel(xlsx_path)

    results = {
        'Depth': {
            'consistent': {'correct': 0, 'total': 0},
            'counter': {'correct': 0, 'total': 0},
            'ambiguous': {'correct': 0, 'total': 0},
        },
        # Distance: excluded — no height-depth entanglement heuristic for inter-object distance
    }

    counter_examples = []

    for _, row in df.iterrows():
        category = row['category']
        if category != 'Depth':
            continue

        hit = row['hit']
        classification, details = classify_cvbench3d_row(row, depth_threshold_ratio)

        if classification not in ['consistent', 'counter', 'ambiguous']:
            continue

        results['Depth'][classification]['total'] += 1
        if hit == 1:
            results['Depth'][classification]['correct'] += 1

        if classification == 'counter':
            counter_examples.append({
                'index': row['index'],
                'category': category,
                'hit': hit,
                'prediction': row['prediction'],
                'answer': row['answer'],
                'source_dataset': row.get('source_dataset', ''),
                'details': details,
            })

    if verbose:
        print("\n=== CV-Bench-3D Depth Classification Statistics ===")
        total = sum(results['Depth'][c]['total'] for c in ['consistent', 'counter', 'ambiguous'])
        print(f"Depth (n={total}):")
        for cls in ['consistent', 'counter', 'ambiguous']:
            n = results['Depth'][cls]['total']
            pct = n / total * 100 if total > 0 else 0
            print(f"  {cls}: {n} ({pct:.1f}%)")
        print("(Distance excluded: no 2D heuristic applies for inter-object 3D distance)")

    return results, counter_examples


# =============================================================================
# Generic report / compare (works for both datasets)
# =============================================================================

_XLSX_SUFFIXES = {
    'embspatial': [
        '_EmbSpatialBench_openai_result',
        '_EmbSpatialBench_exact_matching_result',
    ],
    'cvbench3d': [
        '_CV-Bench-3D_chatgpt-0125_result',
        '_CV-Bench-3D_exact_matching_result',
    ],
}


def extract_model_name(xlsx_path: str, dataset: str) -> str:
    stem = Path(xlsx_path).stem
    for suffix in _XLSX_SUFFIXES.get(dataset, []):
        stem = stem.replace(suffix, '')
    return stem


def print_analysis_report(xlsx_path: str, results: Dict, counter_examples: List[Dict],
                          dataset: str) -> Dict:
    """Print analysis report for a single model (works for any dataset)."""
    model_name = extract_model_name(xlsx_path, dataset)

    print(f"\n{'='*70}")
    print(f"Model: {model_name}")
    print(f"{'='*70}")

    print(f"\n{'Category':<12} {'Type':<12} {'Correct':<10} {'Total':<10} {'Accuracy':<10}")
    print("-" * 54)

    total_consistent = {'correct': 0, 'total': 0}
    total_counter = {'correct': 0, 'total': 0}

    for category in results:
        for cls_type in ['consistent', 'counter', 'ambiguous']:
            data = results[category][cls_type]
            if data['total'] > 0:
                acc = data['correct'] / data['total'] * 100
                print(f"{category:<12} {cls_type:<12} {data['correct']:<10} {data['total']:<10} {acc:.1f}%")

                if cls_type == 'consistent':
                    total_consistent['correct'] += data['correct']
                    total_consistent['total'] += data['total']
                elif cls_type == 'counter':
                    total_counter['correct'] += data['correct']
                    total_counter['total'] += data['total']

    print("-" * 54)
    if total_consistent['total'] > 0:
        acc = total_consistent['correct'] / total_consistent['total'] * 100
        print(f"{'TOTAL':<12} {'consistent':<12} {total_consistent['correct']:<10} {total_consistent['total']:<10} {acc:.1f}%")
    if total_counter['total'] > 0:
        acc = total_counter['correct'] / total_counter['total'] * 100
        print(f"{'TOTAL':<12} {'counter':<12} {total_counter['correct']:<10} {total_counter['total']:<10} {acc:.1f}%")

    if total_consistent['total'] > 0 and total_counter['total'] > 0:
        consistent_acc = total_consistent['correct'] / total_consistent['total'] * 100
        counter_acc = total_counter['correct'] / total_counter['total'] * 100
        gap = consistent_acc - counter_acc
        print(f"\nAccuracy Gap (Consistent - Counter): {gap:.1f}%p")
        print(f"   -> Larger gap indicates stronger reliance on the 2D heuristic")

    counter_wrong = [ex for ex in counter_examples if ex['hit'] == 0]
    if len(counter_wrong) > 0:
        print(f"\n🔍 Counter examples wrong: {len(counter_wrong)} / {len(counter_examples)}")

    return {
        'model_name': model_name,
        'consistent_acc': total_consistent['correct'] / total_consistent['total'] * 100 if total_consistent['total'] > 0 else 0,
        'counter_acc': total_counter['correct'] / total_counter['total'] * 100 if total_counter['total'] > 0 else 0,
        'consistent_total': total_consistent['total'],
        'counter_total': total_counter['total'],
    }


def _run_analysis(xlsx_path: str, dataset: str, cache: Optional[Dict] = None,
                  verbose: bool = False,
                  depth_threshold_ratio: float = 0.05) -> Tuple[Dict, List[Dict]]:
    if dataset == 'cvbench3d':
        return analyze_cvbench3d_results(xlsx_path, verbose=verbose,
                                         depth_threshold_ratio=depth_threshold_ratio)
    else:
        return analyze_embspatial_results(xlsx_path, cache, verbose=verbose)


def compare_models(xlsx_paths: List[str], dataset: str, cache: Optional[Dict] = None):
    """Compare multiple models side by side."""
    summaries = []

    for xlsx_path in xlsx_paths:
        results, counter_examples = _run_analysis(xlsx_path, dataset, cache)
        summary = print_analysis_report(xlsx_path, results, counter_examples, dataset)
        summaries.append(summary)

    max_name_len = max(len(s['model_name']) for s in summaries)
    col_w = max(max_name_len + 2, 40)
    total_w = col_w + 12 + 12 + 10
    print(f"\n{'='*total_w}")
    print("MODEL COMPARISON")
    print(f"{'='*total_w}")
    print(f"{'Model':<{col_w}} {'Consistent':<12} {'Counter':<12} {'Gap':<10}")
    print("-" * total_w)

    for s in summaries:
        gap = s['consistent_acc'] - s['counter_acc']
        print(f"{s['model_name']:<{col_w}} {s['consistent_acc']:.1f}%{'':<6} {s['counter_acc']:.1f}%{'':<6} {gap:+.1f}%p")


EVAL_OUTPUT_DIR = 'VLMEvalKit/outputs'

DEFAULT_MODELS = [
    # Molmo-7B
    'molmo-7B-O-0924/molmo-7B-O-0924',
    'molmo-7B-O-0924-data_scale_exp_80k/molmo-7B-O-0924-data_scale_exp_80k',
    'molmo-7B-O-0924-data_scale_exp_400k/molmo-7B-O-0924-data_scale_exp_400k',
    'molmo-7B-O-0924-data_scale_exp_800k/molmo-7B-O-0924-data_scale_exp_800k',
    'molmo-7B-O-0924-data_scale_exp_2m/molmo-7B-O-0924-data_scale_exp_2m',
    # NVILA-Lite-2B
    'NVILA-Lite-2B/NVILA-Lite-2B',
    'NVILA-Lite-2B-data-scale-exp-80k/NVILA-Lite-2B-data-scale-exp-80k',
    'NVILA-Lite-2B-data-scale-exp-400k/NVILA-Lite-2B-data-scale-exp-400k',
    'NVILA-Lite-2B-data-scale-exp-800k/NVILA-Lite-2B-data-scale-exp-800k',
    'NVILA-Lite-2B-data-scale-exp-2m/NVILA-Lite-2B-data-scale-exp-2m',
    'NVILA-Lite-2B-ST-80k-5pct/NVILA-Lite-2B-ST-80k-5pct',
    'NVILA-Lite-2B-ST-400k-5pct/NVILA-Lite-2B-ST-400k-5pct',
    'NVILA-Lite-2B-ST-800k-5pct/NVILA-Lite-2B-ST-800k-5pct',
    'RoboRefer-2B-SFT/RoboRefer-2B-SFT',
    # Qwen2.5-VL-3B
    'Qwen2.5-VL-3B-Instruct/Qwen2.5-VL-3B-Instruct',
    'Qwen2.5-VL-3B-Instruct-data_scale_exp_80k/Qwen2.5-VL-3B-Instruct-data_scale_exp_80k',
    'Qwen2.5-VL-3B-Instruct-data_scale_exp_400k/Qwen2.5-VL-3B-Instruct-data_scale_exp_400k',
    'Qwen2.5-VL-3B-Instruct-data_scale_exp_800k/Qwen2.5-VL-3B-Instruct-data_scale_exp_800k',
    'Qwen2.5-VL-3B-Instruct-data_scale_exp_2m/Qwen2.5-VL-3B-Instruct-data_scale_exp_2m',
    'Qwen3-VL-235B-A22B-Instruct/Qwen3-VL-235B-A22B-Instruct'
]


def get_default_xlsx_paths(dataset: str) -> List[str]:
    if dataset == 'cvbench3d':
        return [f'{EVAL_OUTPUT_DIR}/{m}_CV-Bench-3D_chatgpt-0125_result.xlsx'
                for m in DEFAULT_MODELS]
    else:
        return [f'{EVAL_OUTPUT_DIR}/{m}_EmbSpatialBench_openai_result.xlsx'
                for m in DEFAULT_MODELS]


def main():
    parser = argparse.ArgumentParser(description='Counter vs Consistent Example Analysis')
    parser.add_argument('xlsx_files', nargs='*',
                        help='Model result xlsx files (uses default model list if omitted)')
    parser.add_argument('--dataset', choices=['embspatial', 'cvbench3d'], default='embspatial',
                        help='Benchmark dataset to analyze (default: embspatial)')
    parser.add_argument('--compare', action='store_true', help='Compare multiple models')
    parser.add_argument('--verbose', '-v', action='store_true', help='Verbose output')
    parser.add_argument('--output', '-o', type=str, help='Save results to file')
    parser.add_argument('--save-cache', type=str,
                        help='Save EmbSpatialBench classification cache to JSON')
    parser.add_argument('--load-cache', type=str,
                        help='Load EmbSpatialBench classification cache from JSON')

    args = parser.parse_args()

    # Build/load cache (EmbSpatialBench only; CV-Bench-3D reads bbox from xlsx directly)
    cache = None
    if args.dataset == 'embspatial':
        if args.load_cache and Path(args.load_cache).exists():
            print(f"Loading cache from {args.load_cache}...")
            with open(args.load_cache, 'r') as f:
                cache = json.load(f)
        else:
            cache = build_classification_cache(verbose=args.verbose)

        if args.save_cache:
            print(f"Saving cache to {args.save_cache}...")
            with open(args.save_cache, 'w') as f:
                json.dump(cache, f, indent=2)

    xlsx_files = args.xlsx_files if args.xlsx_files else get_default_xlsx_paths(args.dataset)

    tee = None
    if args.output:
        tee = TeeWriter(args.output)
        sys.stdout = tee

    try:
        if args.compare or len(xlsx_files) > 1:
            compare_models(xlsx_files, args.dataset, cache)
        else:
            results, counter_examples = _run_analysis(
                xlsx_files[0], args.dataset, cache, args.verbose
            )
            print_analysis_report(xlsx_files[0], results, counter_examples, args.dataset)
    finally:
        if tee is not None:
            sys.stdout = tee.close()
            print(f"Results saved to {args.output}")


if __name__ == '__main__':
    main()