seujung / SNAIL-gluon

Implementation of SNAIL(A Simple Neural Attentive Meta-Learner) with Gluon

SNAIL with Gluon

---

Gluon inplementation of A Simple Neural Attentive Meta-Learniner

network structore

building block structure

Requirements

• Python 3.6.1
• mxnet 1.3.1
• mxboard 0.1.0
• tqdm 4.29.0

Application

• Omniglot

Usage

• arguments
  • batch_size : Define batch size (defualt=64)
  • epochs : Define total epoches (default=50)
  • N : the nunber of N-way (default=10)
  • K : the number of K-shot (default=5)
  • iterations : the number of data iteration (default=1000)
  • input_dims : embedding dimension of input data (default=64)
  • download : download omniglot dataset (default=False)
  • GPU_COUNT : use gpu count (default=1)
  • logdir : location of mxboard log file (default=./log)
  • modeldir : location of model parameter file (default=./models)

default setting

python main.py

or

manual setting

python main.py --batch_size=24 --epochs=200 ..

Results

10-way 5-shot case

Reference

• https://github.com/sagelywizard/snail
• https://github.com/eambutu/snail-pytorch