import json import os from datasets import Dataset, Image from transformers import CLIPProcessor, CLIPModel, AutoTokenizer, AutoModel from PIL import Image as PILImage import torch from tqdm import tqdm def preprocess_multimodal_data(raw_json_path, image_dir, dataset_name, text_model_name, clip_model_name, output_dir): """ Preprocesses raw JSON and image data into a Hugging Face Dataset, using DeepSeek-R1 for text embeddings. Args: raw_json_path (str): Path to the raw JSON file. image_dir (str): Directory containing the images. dataset_name (str): Name of the Hugging Face Dataset. text_model_name (str): Name of the text model (DeepSeek-R1) from Hugging Face Transformers. clip_model_name (str): Name of the CLIP model from Hugging Face Transformers. output_dir (str): Directory where the dataset repository will be saved. """ # Load data from JSON with open(raw_json_path, 'r') as f: data = json.load(f) # Initialize lists to store data images = [] questions = [] answers = [] ids = [] subjects = [] topics = [] options = [] certification_levels = [] difficulty_levels = [] grade_levels = [] learning_objectives = [] question_types = [] time_requireds = [] keywords_list = [] sources = [] years = [] sections = [] languages = [] # Load CLIP, DeepSeek-R1, and Tokenizer clip_processor = CLIPProcessor.from_pretrained(clip_model_name) clip_model = CLIPModel.from_pretrained(clip_model_name).to("cuda") text_model = AutoModel.from_pretrained(text_model_name).to("cuda") tokenizer = AutoTokenizer.from_pretrained(text_model_name) # Process each data point for i, item in enumerate(tqdm(data, desc="Processing data")): question = item.get('question', '') answer = item.get('answer', '') image = None image_path = item.get('image', None) id_field = str(i) if image_path: try: full_image_path = os.path.join(image_dir, image_path) image = PILImage.open(full_image_path).convert("RGB") except FileNotFoundError: print(f"Warning: Image not found at {full_image_path}") except Exception as e: print(f"Warning: Error processing image {full_image_path}: {e}") images.append(image) questions.append(question) answers.append(answer) ids.append(id_field) subjects.append(item.get('subject', None)) topics.append(item.get('topic', None)) options.append(item.get('options', None)) certification_levels.append(item.get('certification_level', None)) difficulty_levels.append(item.get('difficulty_level', None)) grade_levels.append(item.get('grade_level', None)) learning_objectives.append(item.get('learning_objective', None)) question_types.append(item.get('question_type', None)) time_requireds.append(item.get('time_required', None)) keywords_list.append(item.get('keywords', None)) sources.append(item.get('source', None)) years.append(item.get('year', None)) sections.append(item.get('section', None)) languages.append(item.get('language', None)) # Create Hugging Face Dataset dataset_dict = { 'image': images, 'question': questions, 'answer': answers, 'id': ids, 'subject': subjects, 'topic': topics, 'options': options, 'certification_level': certification_levels, 'difficulty_level': difficulty_levels, 'grade_level': grade_levels, 'learning_objective': learning_objectives, 'question_type': question_types, 'time_required': time_requireds, 'keywords': keywords_list, 'source': sources, 'year': years, 'section': sections, 'language': languages, } dataset = Dataset.from_dict(dataset_dict) dataset = dataset.cast_column("image", Image()) # Save to Hugging Face Dataset repository dataset.save_to_disk(os.path.join(output_dir, dataset_name)) print(f"Dataset '{dataset_name}' saved to {os.path.join(output_dir, dataset_name)}") # Example Usage: raw_json_path = 'your_data.json' image_dir = 'your_image_dir' dataset_name = 'my_multimodal_dataset' text_model_name = 'deepseek-ai/deepseek-rl-7b-chat' #Deepseek R1 clip_model_name = 'openai/clip-vit-base-patch32' output_dir = 'my_datasets' preprocess_multimodal_data(raw_json_path, image_dir, dataset_name, text_model_name, clip_model_name, output_dir)