eval.py

import os
import json
import torch
import torch.multiprocessing as mp
import argparse
import numpy as np
import random
from collections import defaultdict
from transformers import Qwen3VLForConditionalGeneration, AutoProcessor
from qwen_vl_utils import process_vision_info
from tqdm import tqdm
import threading
import time

CONFIG = {
    "MODEL_PATH": "path/to/Qwen3-VL-8B-Instruct",
    "TARGET_JSON_PATH": "path/to/test.json",
    "max_pixels": 512**2,
    "DEFAULT_GPUS": 8,
    "max_tokens": 128,
}

def load_model_and_processor(model_path, device):
    model = Qwen3VLForConditionalGeneration.from_pretrained(
        model_path,
        dtype="auto"
    ).to(device)
    processor = AutoProcessor.from_pretrained(model_path)
    return model, processor

def build_interleaved_messages(qa_item):
    content_list = []
    videos_data = qa_item.get("videos", [])
    question_text = qa_item.get("question", "")

    if isinstance(videos_data, list):
        num_video_placeholders = question_text.count('<video>')
        
        if num_video_placeholders == len(videos_data):
            text_parts = question_text.split('<video>')
            for i, video_data in enumerate(videos_data):
                if text_parts[i]:
                    content_list.append({"type": "text", "text": text_parts[i]})
                content_list.append(
                    {
                        "type": "video", 
                        "video": video_data['video'],
                        "sample_fps": video_data['sample_fps'],
                        "max_pixels":CONFIG["max_pixels"],
                        "min_pixels":128 ** 2
                    })
            if text_parts[-1]:
                content_list.append({"type": "text", "text": text_parts[-1]})
        else:
            print("build_interleaved_messages_error")
        
    return [{
            "content": "Based on this video, provide the answer directly.",
            "role": "system"
        },
        {
            "role": "user", 
            "content": content_list
        }]


def calculate_score(prediction, ground_truth, task_type):
    if task_type == "Jigsaw":
        return 1.0 if str(prediction).strip() == str(ground_truth).strip() else 0.0

    elif task_type == "Counting":
        try:
            gt_parts = [p.strip() for p in str(ground_truth).split(',')]
            pred_parts = [p.strip() for p in str(prediction).split(',')]
            
            if len(gt_parts) != 3:
                print(f"Warning: Ground truth for Counting task '{ground_truth}' does not have 3 parts.")
                return 0.0

            total_score = 0.0

            for i in range(3):
                if i < len(pred_parts) and pred_parts[i] == gt_parts[i]:
                    total_score += 1.0
            
            return total_score / 3.0
        except Exception as e:
            print(f"Error parsing Counting prediction: {e}")
            return 0.0

    elif task_type == "Grounding": 
        try:
            gt_start, gt_end = map(float, str(ground_truth).split('-'))
            pred_start, pred_end = map(float, str(prediction).split('-'))

            inter_start = max(gt_start, pred_start)
            inter_end = min(gt_end, pred_end)

            intersection_area = max(0, inter_end - inter_start)

            gt_area = gt_end - gt_start
            pred_area = pred_end - pred_start
            union_area = gt_area + pred_area - intersection_area

            if union_area == 0:
                return 1.0 if intersection_area > 0 else 0.0
            
            iou = intersection_area / union_area
            return iou
        except (ValueError, IndexError) as e:
            return 0.0
            

def process_single_item(qa_item, model, processor):
    correct_answer = qa_item.get("answer")
    ability_type = qa_item.get("ability_type")
    task_type = qa_item.get("task_type")
    subset = qa_item.get("subset")

    try:
        messages = build_interleaved_messages(qa_item)
        text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
        images, videos, video_kwargs = process_vision_info(messages, image_patch_size=16, return_video_kwargs=True, return_video_metadata=True)
        if videos is not None:
            videos, video_metadatas = zip(*videos)
            videos, video_metadatas = list(videos), list(video_metadatas)
        else:
            video_metadatas = None

        inputs = processor(text=text, images=images, videos=videos, video_metadata=video_metadatas, return_tensors="pt", do_resize=False, **video_kwargs)
        inputs = inputs.to(model.device)
        
        generated_ids = model.generate(**inputs, max_new_tokens=CONFIG["max_tokens"], do_sample=False)
        generated_ids_trimmed = [out_ids[len(in_ids):] for in_ids, out_ids in zip(inputs.input_ids, generated_ids)]
        output_text = processor.batch_decode(
            generated_ids_trimmed, skip_special_tokens=True, clean_up_tokenization_spaces=False
        )[0]
        predicted_answer = output_text 
        
        score = calculate_score(predicted_answer, correct_answer, task_type)
        
        print(f"Task: {task_type} | Predicted: {predicted_answer} | Correct: {correct_answer} | Score: {score:.2f}")

        return ("success", score, ability_type, task_type, subset)

    except Exception as e:
        print(f"Error processing item for subset {subset}: {e}")
        return ("error", 0.0, ability_type, task_type, subset)


def aggregate_and_print_results(all_results, total_dataset_size):
    if not all_results:
        print("No results to aggregate.")
        return

    scores_by_ability = defaultdict(list)
    scores_by_task = defaultdict(list)
    scores_by_subset = defaultdict(list) 
    error_count = 0
    success_count = 0

    for result in all_results:
        status, score, ability, task, subset = result
        scores_by_ability[ability].append(score)
        scores_by_task[task].append(score)
        scores_by_subset[subset].append(score) 

        if status == "success":
            success_count += 1
        else:
            error_count += 1

    def print_report(title, score_dict):
        print(f"--- {title.upper()} AVERAGE SCORES ---")
        if not score_dict:
            print("No data available.")
            return
            
        total_scores = []
        for name, scores in sorted(score_dict.items()):
            avg_score = np.mean(scores) if scores else 0
            total_scores.extend(scores)
            print(f"{name:<25} | Average Score: {avg_score:.4f} (N={len(scores)})")
        
        overall_avg = np.mean(total_scores) if total_scores else 0
        print("-" * 50)
        print(f"{'Overall ' + title:<25} | Average Score: {overall_avg:.4f} (N={len(total_scores)})")
        print("-" * 50)


    print("\n" + "="*50)
    print(" " * 16 + "EVALUATION RESULTS")
    print("="*50)
    print(f"Model Path: {CONFIG['MODEL_PATH']}")
    print(f"Total Items in Dataset: {total_dataset_size}")
    print(f"Successfully Processed: {success_count}")
    print(f"Processing Errors: {error_count}\n")

    print_report("Ability", scores_by_ability)
    print("")
    print_report("Task", scores_by_task)
    print("")
    print_report("Subset", scores_by_subset) 
    print("="*50)

def worker(rank, task_queue, results_list):
    os.environ["CUDA_VISIBLE_DEVICES"] = str(rank)
    device = torch.device(f"cuda:{str(rank)}")

    model, processor = load_model_and_processor(CONFIG["MODEL_PATH"], device)

    while True:
        try:
            qa_item = task_queue.get_nowait()
        except Exception:
            break 

        result = process_single_item(qa_item, model, processor)
        results_list.append(result)

def progress_monitor(results_list, total_size, stop_event):
    with tqdm(total=total_size, desc="Processing Items", unit="item") as pbar:
        while not stop_event.is_set():
            current_count = len(results_list)
            pbar.update(current_count - pbar.n)
            if current_count >= total_size:
                break
            time.sleep(1)
        pbar.update(total_size - pbar.n)


def main():
    world_size = CONFIG['DEFAULT_GPUS']

    try:
        with open(CONFIG["TARGET_JSON_PATH"], 'r', encoding='utf-8') as f:
            full_dataset = json.load(f)
        total_dataset_size = len(full_dataset)
        print(f"Successfully loaded {total_dataset_size} items from {CONFIG['TARGET_JSON_PATH']}.")
    except Exception as e:
        print(f"Fatal: Error loading dataset. {e}")
        return

    random.shuffle(full_dataset)

    with mp.Manager() as manager:
        task_queue = manager.Queue()
        for item in full_dataset:
            task_queue.put(item)

        results_list = manager.list()

        stop_event = threading.Event()
        progress_thread = threading.Thread(
            target=progress_monitor,
            args=(results_list, total_dataset_size, stop_event)
        )
        progress_thread.start()

        mp.spawn(worker,
                 args=(task_queue, results_list),
                 nprocs=world_size,
                 join=True)

        stop_event.set()
        progress_thread.join()

        final_results = list(results_list)

    aggregate_and_print_results(final_results, total_dataset_size)


if __name__ == "__main__":
    mp.set_start_method("spawn", force=True)
    main()