yuwei-z / Image-Caption

project for machine learning

Image-Caption

Project for IDS 705: Principles of Machine Learning

Contents

• Image-Caption
  • References
  • Before training
  • Preprocess
  • Training Process
  • Evaluation metrics
  • Testing Process

---

References

Reference for network framework and training code: a-PyTorch-Tutorial-to-Image-Captioning.
Reference Paper Show, Attend and Tell
Data comes from: Flickr 8k Dataset

---

Before training

1. clone the current repository to servers

$ git clone [e-mail]:wkhalil/Image-Caption.git

2. before training & testing: make sure all libraries are installed with compatible versions (requirements.txt for reference).

• using the interpreter with compatible environment to open the project.
• if using MacOS Catalina, please run the following commands to avoid potential bug. https://stackoverflow.com/questions/48290403/process-finished-with-exit-code-134-interrupted-by-signal-6-sigabrt

3. check for directory tree (missing folders can be created manually, most with place_holder.txt)
   download Flickr 8k Dataset and store the folder with images in project as './inputs/Images'.
   directory necessary before training:

.
├── README.md
├── caption.py
├── config.py
├── create_input_files.py
├── datasets.py
├── eval.py
├── inputs
│   ├── Images
│   ├── Intermediate_files
│   │   └── place_holder.txt
│   └── captions.txt
├── logs
│   └── log
│       ├── train
│       │   └── place_holder.txt
│       └── val
│           └── place_holder.txt
├── model_history
│   └── place_holder.txt
├── models.py
├── output
│   └── place_holder.txt
├── requirements.txt
├── test_results
│   └── place_holder.txt
├── train.py
└── utils.py

---

Preprocess

$ python create_input_files.py

(only once for generating intermediate data files)

Training Process

1. There are two choices for training, if training without previous checkpoints, set checkpoints=None in config.py.
   Another one is training based on current best model (default, obtain best model: latest model checkpoints).
2. run the train.py

$ python train.py

---

Evaluation metrics

first install pycocoevalcap for CIDER, SPICE metrics.(Problems still exist for SPICE after installing following instructions, and others also met the same problem)

$ python eval.py

---

Testing Process

All generated captions would be stored under test_results directory.
Three choices:

1. If randomly generated one caption from all inputs

$ python caption.py

2. If randomly generated multiple captions from all inputs
   ex. randomly select 6 images to generate captions

$ python caption.py --num = 6

3. generate caption for specified image & specified model & specified path to save
   ex.

$ python caption.py  --word_map='./inputs/Intermediate_files/WORDMAP_flickr8k_5_cap_per_img_5_min_word_freq.json' --beam_size=5 --img='./inputs/Images/2877424957_9beb1dc49a.jpg' --save='./test_results/gen_3' --model='./model_history/best_0417_20.pth.tar'