cjf8899 / Tiny_ImageNet_Challenge

:clipboard:This is a miniature of ImageNet classification Challenge.

Tiny_ImageNet_Challenge

Tiny ImageNet Challenge: This is a miniature of ImageNet classification Challenge.

• Docker >= 19.03.8
• CUDA >= 10.2

Getting Started

Download git and dataset

git clone https://github.com/cjf8899/Tiny_ImageNet_Challenge.git

cd Tiny_ImageNet_Challenge
cd data
sh download_and_unzip.sh

# Pretrain model download(ImageNet)
wget http://data.lip6.fr/cadene/pretrainedmodels/se_resnext50_32x4d-a260b3a4.pth

the structures would like

~/Tiny_ImageNet_Challenge/data/
    -- tiny-imagenet-200
    -- se_resnext50_32x4d-a260b3a4.pth

Train

I used wandb and various other transforms as well.

The wandb code is included in this repositories.

Wandb Guide

All of the special transforms I used are included in this repositories.

Results

I used SE-ResNeXt50_32x4d and the best performance is 82.54%

Implementation	top-1	top-5
SE-ResNeXt50	82.54	94.96