Skip to content

README.md

Latest commit

 

History

History
executable file
·
54 lines (37 loc) · 1.81 KB

README.md

File metadata and controls

executable file
·
54 lines (37 loc) · 1.81 KB

Image Classification with Qwen 2.5 VL-32B Instruct

Automated large-scale image classification using Qwen/Qwen2.5-VL-32B-Instruct vision model. Processes images in parallel and outputs results to CSV files.

Features

  • Parallel batch processing using configurable thread pool (32 workers)
  • Checkpointing: saves progress every 100 images or 300 seconds
  • Detailed logging of successes, failures, and token usage
  • 26 Categories: Product, Food, Finance, Architecture, Sport, etc.

Prerequisites

  • Python 3.8+
  • OpenAI Python client: pip install openai
  • Valid API key for Qwen model access

Setup

  1. Set API Key:

    export DASHSCOPE_API_KEY="your-api-key-here"

  2. Change BASE_URL

    • Update the BASE_URL variable in the script.

  3. Prepare Image Directory:

    • Place images in ../dataset/autodl-tmp/60k_jpg/
    • Supported formats: .jpg, .jpeg, .png

  4. Configure Script (optional):

    • Edit IMAGE_DIR, OUTPUT_CSV, FAILED_CSV paths
    • Adjust MAX_WORKERS and save intervals

Usage

python image_classification_60k.py

Output

  • Success: ../dataset/60k_image_classification.csv
    • Columns: image_name, category, prompt_tokens, completion_tokens, total_tokens
  • Failures: ../dataset/60k_failed_image_classification.csv
    • Columns: image_name, failure_reason
  • Logs: classification.log

Categories

The model classifies images into 26 categories:

  1. Product, 2. Food & Beverage, 3. Finance, 4. Architecture, 5. Sport, 6. Transport, 7. Art, 8. Urban, 9. Nature, 10. People, 11. Event, 12. History & Culture, 13. Animation & Comics, 14. Science & Technology, 15. Statistical Data, 16. Poster, 17. Interior, 18. Game, 19. Snapshot, 20. Education, 21. News & Newspaper, 22. Tourism, 23. Advertisement, 24. Meme, 25. UI, 26. Others