Skip to content

FashionAI Global Challenge—Attributes Recognition of Apparel—Ranked 10st solution.

Notifications You must be signed in to change notification settings

gd2016229035/FashionAI2018-TianChi

Repository files navigation

FashionAI2018-TianChi

Introduction

This is the main Gluon code of 阿里天池竞赛——服饰属性标签识别. Note that this code is just a part of our final code, but provides the one of our best single model. Final submission is a ensemble model of two model: One is this resnet152v2 model ,and the other is the Inceptionv4 model from my teammates. The code is based on hetong007's code which provides a good baseline in the competition. This is my first time to use Gluon and thanks to hetong007~

Team name:时尚时尚最时尚 Rank: 10/2950 (Season1) 17/2950 (Season2)

Software:

  • ubuntu14.04,cuda8.0,cudnn6.5
  • python2.7
  • mxnet-cu80
  • numpy
  • pandas

Highlights

  • Higher performance: Improve the result by many modifications for a pure single model(without backbone ensemble).
  • Faster training speed: Use gluon.data.vision.transforms for data augmentation which is faster than original code.
  • Soft label: Define our own cutom dataset and treat 'maybe'(m) label as 'soft label' when training which can boost the result.
  • Muti-scale train & test: Use more scale augmentations when training and testing, especially for TTA(Test time augmentation).
  • mAP defination: Define mAP by ourselves according the competetion illustrate.
  • Random erasing: Gluon version code defined by ourselvers.
  • Heatmap strategy: Find the circumscribed square of the largest connected block, framing the entire heat map area for classification finetuning.

Training in a few lines

  1. Download data and extract it into data1/(season1) and data2/(season2).
  2. python2 prepare_data.py Prepare dataset for trainset, valset and testset.
  3. bash benchmark.sh
  • num_gpus,set to 1 for single GPU training

About

FashionAI Global Challenge—Attributes Recognition of Apparel—Ranked 10st solution.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published