Deep Low-Shot Learning for Biological Image Classification and Visualization from Small Training Samples

Created by Lei Cai, and Shuiwang Ji.

Introduction

Deep low-shot learning for precise stage annotation for embryo images. BDGP dataset contains 36908 images with stage range label in which 3474 images are give precise stage label.

We employ deep low-shot learning to annotate the embryo images automatically.

System requirement

Programming language

Python 3.5+

Python Packages

Pytorch, numpy

Training

Train the residual network

python main.py --mode=0

Train the residual network with transfer learning

python main.py --mode=1

Generate reference set

python main.py --mode=2

Fine-tune the network

python fintune.py

GEM results