DATASET.md

Datasets for few-shot learning

Simply execute:

sh dataset/get_tier_and_mini.sh

You are all set!

---

The following are just some documents to get you more comfortable about datasets.

• miniImagenet

  • Thrid-party instructions (OpenAI's Reptile) here (update: do NOT use it as it is too large; it will download the original large images from ImageNet). Google drive file here to directly download the mini-imagenet.zip file (we use this).

  • The splits test.csv, train.csv, val.csv (already there if you clone our repo) can be downloaded from Ravi and Larochelle - splits. For more information on how to obtain the images check the original source Ravi and Larochelle - github.

• tierImagenet

  • Original info here.

Structure

The dataset folder looks like this:

few-shot-ctm
├── ...
└── dataset
   |__ data_loader.py
   |__ ...
   
   |__ miniImagenet                
      └── images            # (ignored in repo)
         ├── n0153282900000006.jpg
         ├── n1313361300001299.jpg
         └── ...
      |__ tes.csv
      |__ train.csv
      |__ val.csv
      
   |__ tier_split
      |__ train.csv
      ...
      
   |__ tier_imagenet        # (ignored in repo)
      |__ train_images_png.pkl 
      |__ train_labels.pkl
      |__ ...