CLIP (Contrastive Language–Image Pre-training)

Experiments (Evaluation)

Overview

Training

• Work In Process

Usage

• Evaluation

python evaluation.py --dataset CIFAR100 --cuda True

• args
  • dataset (str): CIFAR10, CIFAR100 (default: CIFAR100)
  • num_workers (int): default: 0
  • batch_size (int): default: 128
  • cuda (bool): False
• Training
  • Prepare Data
    • Visual Genome Dataset link
    • Download (images, region descriptions)
  • training

  python main.py --base_dir ./ --cuda True
  

Reference

• paper link
• Author: Alec Radford, Jong Wook Kim, Chris Hallacy, Girish Sastry, Amanda Askell, Pamela Mishkin, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Jack Clark, Gretchen Krueger, Ilya Sutskever
• OpenAI