run_image_model.py

import os, sys
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
try:
    from vlmeval.config import supported_VLM
except Exception:  # only needed for local ("image") models
    supported_VLM = None
import json
import re
import base64
import numpy as np
import torch
import time
try:
    import colorama
except Exception:
    colorama = None
from util.prompt_generation import generate_question_prompt
try:
    from colorama import Fore, Style, init
except Exception:  # colorama is optional (cosmetic terminal colors)
    class _NoColor:
        def __getattr__(self, _):
            return ""
    Fore = Style = _NoColor()
    def init(*a, **k):
        return None
import cv2
from pathlib import Path
from util.concate_image import concatenate_image
from PIL import Image
from io import BytesIO
import tempfile
import argparse
import os

SAMPLE_FRAMES = 16

def fix_seed():
    import random
    random.seed(42)
    np.random.seed(42)
    torch.manual_seed(42)
    if torch.cuda.is_available():
        torch.cuda.manual_seed_all(42)

def concate_image_to_video(images, img_path):
    """
    Appends an external image (from img_path) to the list of PIL images.
    """
    extra_img = Image.open(img_path).convert("RGB")  # open and ensure RGB
    images.append(extra_img)
    return images

def generate_input_images_memory(mp4_path, num_frames=SAMPLE_FRAMES):
    """
    Extracts `num_frames` frames from a video and returns them as PIL.Image objects in memory.
    """
    cap = cv2.VideoCapture(mp4_path)
    if not cap.isOpened():
        raise ValueError(f"Cannot open video file: {mp4_path}")

    total_frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
    frame_indices = np.linspace(0, total_frames - 1, num=num_frames, dtype=int)

    images = []
    for i in frame_indices:
        cap.set(cv2.CAP_PROP_POS_FRAMES, i)
        success, frame = cap.read()
        if not success:
            continue

        # Convert OpenCV BGR image to RGB, then to PIL.Image
        rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
        image = Image.fromarray(rgb)
        images.append(image)

    cap.release()
    return images

def evaluate_model_video(category_name, model_name, item):
    
    fix_seed()

    item_copy = item.copy()
    
    question = item.get("question", "")
    options = item.get("options", [])
    video = item.get("video", None)

    for format in ["direct"]:
        question_prompt = generate_question_prompt(format, category_name, question, options, model_category="image")
        print(f"{Fore.YELLOW}Evaluating item{Style.RESET_ALL}")
        print(f"{Fore.BLUE}Question Prompt: {question_prompt}{Style.RESET_ALL}")
        # print the input question prompt
        random_number = np.random.randint(0, 9999)
        input_images_open = generate_input_images_memory(video, SAMPLE_FRAMES)
        # conbine the input_images

        # debug this function
        combined_image = concatenate_image(input_images_open, rows=2, columns=8)

        # Use tempfile to avoid collisions between process
        with tempfile.NamedTemporaryFile(suffix=".jpg", prefix="temp_combined_image_", delete=False) as tmp_file:
            temp_image_path = tmp_file.name
            combined_image.save(temp_image_path, format="JPEG")

        print(f"{Fore.GREEN}Combined image saved as {temp_image_path}{Style.RESET_ALL}")

        response = supported_VLM[model_name]().generate([temp_image_path, question_prompt])

        # Optional: clean up after inference
        Path(temp_image_path).unlink(missing_ok=True)


        print(f"{Fore.CYAN}Response is: {response}{Style.RESET_ALL}")
        item_copy[f"{format}_reply"] = response
    
    return item_copy

def evaluate_model_interleaved(category_name, model_name, item):
    
    fix_seed()
    # load input_json file

    item_copy = item.copy()
        
    question = item.get("question", "")
    options = item.get("options", [])
    video = item.get("video", None)
    image = item.get("image", None)

    for format in ["direct"]:
        question_prompt = generate_question_prompt(format, category_name, question, options, model_category="image")
        print(f"{Fore.YELLOW}Evaluating item: {Style.RESET_ALL}")
        print(f"{Fore.BLUE}Question Prompt: {question_prompt}{Style.RESET_ALL}")
        # print the input question prompt
        random_number = np.random.randint(0, 9999)
        input_images_open = generate_input_images_memory(video, SAMPLE_FRAMES)
        # conbine the input_images
        input_images_open_new = concate_image_to_video(input_images_open, image)

        # debug this function
        combined_image = concatenate_image(input_images_open_new, rows=5, columns=7)

        # Use tempfile to avoid collisions between process
        with tempfile.NamedTemporaryFile(suffix=".jpg", prefix="temp_combined_image_", delete=False) as tmp_file:
            temp_image_path = tmp_file.name
            combined_image.save(temp_image_path, format="JPEG")

        print(f"{Fore.GREEN}Combined image saved as {temp_image_path}{Style.RESET_ALL}")

        response = supported_VLM[model_name]().generate([temp_image_path, question_prompt])

        # Optional: clean up after inference
        Path(temp_image_path).unlink(missing_ok=True)

        print(f"{Fore.YELLOW}Response for item: {response}{Style.RESET_ALL}")
        item_copy[f"{format}_reply"] = response

    return item_copy

def evaluate_model_image(category_name, model_name, item):
    
    fix_seed()

    item_copy = item.copy()

    question = item.get("question", "")
    options = item.get("options", [])
    image = item.get("image", None)

    for format in ["direct"]:
        question_prompt = generate_question_prompt(format, category_name, question, options, model_category="image")
        print(f"{Fore.YELLOW}Evaluating item: {Style.RESET_ALL}")
        # print the input question prompt

        # if failure, try a few more attempts
        response = "error"
        for attempt in range(3):
            try:
                response = supported_VLM[model_name]().generate([image, question_prompt])
                break
            except:
                print(f"{Fore.RED}Error generating response for item, attempt {attempt + 1}. Retrying...{Style.RESET_ALL}")
                time.sleep(2 ** attempt)
                continue  # retry the same item

        item_copy[f"{format}_reply"] = response

    return item_copy

if __name__ == "__main__":
    # only evaluate model within the certain category in the input json file
    
    parser = argparse.ArgumentParser(description="Evaluate VLM model on video QA tasks.")
    parser.add_argument("--model_name", type=str, required=True, help="The specific model name (e.g., claude-sonnet-4-20250514)")
    parser.add_argument("--input_json_path", type=str, required=True, help="Path to the input JSON file.")
    parser.add_argument("--category_name", type=str, required=False, default="", help="Optional. Routing is auto-detected from each item's category.")
    parser.add_argument("--model_series", type=str, required=False, default="image", help="Unused for local image models; kept for CLI symmetry.")
    parser.add_argument("--evaluate_output_category", type=str, default=None, help="Path to save evaluation results. Default is auto-generated.")

    args = parser.parse_args()
    
    model_name = args.model_name
    input_json_path = args.input_json_path
    category_name = args.category_name
    model_series = args.model_series
    evaluate_category = args.evaluate_output_category

    print(f"{Fore.GREEN}Evaluating model: {model_name} on category: {category_name}{Style.RESET_ALL}")

    evaluate_output_path =f"final_{model_name}_evaluate_{evaluate_category}.json"
    tmp_dir = "tmp"
    os.makedirs(tmp_dir, exist_ok=True)
    tmp_path = os.path.join(tmp_dir, evaluate_output_path)

    # Load input
    with open(input_json_path, "r", encoding="utf-8") as f:
        input_data = json.load(f)

    # === RESUME LOGIC ===
    evaluate_output = []
    processed_ids = set()

    if os.path.exists(tmp_path):
        try:
            with open(tmp_path, "r", encoding="utf-8") as f:
                evaluate_output = json.load(f)
            # Build fast lookup of processed items: prefer explicit idx, else ordinal
            for pos, item in enumerate(evaluate_output):
                key = item.get("idx", pos)
                processed_ids.add(key)
            print(f"{Fore.YELLOW}[Resume] Loaded {len(evaluate_output)} partial results from {tmp_path}.{Style.RESET_ALL}")
        except Exception as e:
            print(f"{Fore.RED}[Resume] Could not load {tmp_path}: {e}. Starting fresh.{Style.RESET_ALL}")
            evaluate_output = []
            processed_ids = set()

    # Iterate and skip already processed
    for pos, item in enumerate(input_data):
        key = item.get("idx", pos)
        if key in processed_ids:
            continue  # already done

        subcat = item.get("category", "")
        if subcat in ["pointing", "trajectory", "bbox"]:
            output_item = evaluate_model_image(subcat, model_name, item)
        elif subcat in ["path planning", "relative direction"]:
            output_item = evaluate_model_interleaved(subcat, model_name, item)
        elif subcat in ["object localization", "next action prediction", "task progress reasoning"]:
            output_item = evaluate_model_video(subcat, model_name, item)
        elif item.get("video"):
            output_item = evaluate_model_video(subcat, model_name, item)
        elif item.get("image"):
            output_item = evaluate_model_image(subcat, model_name, item)
        else:
            output_item = item

        evaluate_output.append(output_item)
        processed_ids.add(key)

        # Save at tmp
        with open(tmp_path, "w", encoding="utf-8") as f:
            json.dump(evaluate_output, f, indent=4, ensure_ascii=False)

    # Write final output
    with open(evaluate_output_path, "w", encoding="utf-8") as f:
        json.dump(evaluate_output, f, indent=4, ensure_ascii=False)

    print(f"{Fore.CYAN} Done. Saved final results to: {evaluate_output_path}{Style.RESET_ALL}")