Tools for mini-ImageNet Dataset

This repo provides python source codes for creating mini-ImageNet dataset from ImageNet and the utils for generating batches during training.

Summary

• About mini-ImageNet
• Requirements
• Usage
• Performance
• Acknowledgement

About mini-ImageNet

Mini-ImageNet dataset was proposed by Vinyals et al. for few-shot learning evaluation. Its complexity is high due to the use of ImageNet images but requires fewer resources and infrastructure than running on the full ImageNet dataset. In total, there are 100 classes with 600 samples of 84×84 color images per class. These 100 classes are divided into 64, 16, and 20 classes respectively for sampling tasks for meta-training, meta-validation, and meta-test.

Requirements

• Python 2.7 or 3.x
• numpy
• tqdm
• opencv-python
• Pillow

Usage

First, you need to download the image source files from ImageNet website. If you already have it, you may use it directly.

Filename: ILSVRC2012_img_train.tar
Size: 138 GB
MD5: 1d675b47d978889d74fa0da5fadfb00e

Then clone the repo:

git clone https://github.com:y2l/mini-imagenet-tools.git
cd mini-imagenet-tools

To generate mini-ImageNet dataset from tar file:

python mini_imagenet_generator.py --tar_dir [your_path_of_the_ILSVRC2012_img_train.tar]

To generate mini-ImageNet dataset from untarred folder:

python mini_imagenet_generator.py --imagenet_dir [your_path_of_imagenet_folder]

If you want to resize the images to the specified resolution:

python mini_imagenet_generator.py --tar_dir [your_path_of_the_ILSVRC2012_img_train.tar] --image_resize 100

P.S. In default settings, the images will be resized to 84 × 84.

If you don't want to resize the images, you may set --image_resize 0.

To use the MiniImageNetDataLoader class:

from mini_imagenet_dataloader import MiniImageNetDataLoader

dataloader = MiniImageNetDataLoader(shot_num=5, way_num=5, episode_test_sample_num=15)

dataloader.generate_data_list(phase='train')
dataloader.generate_data_list(phase='val')
dataloader.generate_data_list(phase='test')

dataloader.load_list(phase='all')

for idx in range(total_train_step):
    episode_train_img, episode_train_label, episode_test_img, episode_test_label = \
        dataloader.get_batch(phase='train', idx=idx)
    ...

Performance

The 1-shot, 5-way classification accuracy (%)

Method	Accuracy
MAML	48.70 ± 1.75
ProtoNets	49.42 ± 0.78
SNAIL	55.71 ± 0.99
TADAM	58.5 ± 0.3
MTL	61.2 ± 1.8

Acknowledgement

MAML

Optimization as a Model for Few-Shot Learning