aihgf / fashionai-keypoints-detection-pytorch

[WIP] A refactored version of https://github.com/gathierry/FashionAI-KeyPointsDetectionOfApparel

Introduction

The code is originally from Shiyu's GitHub repository (gathierry/FashionAI-KeyPointsDetectionOfApparel/), which provided his solution that achieved LB 3.82% in Tianchi FashionAI Global Challenge, 17th place out 2322 teams. For educational purpose, I've refactored most of his code and added plenty of documentations to help me understand what's happening behind the scene.

I could have made some mistakes along the way. For the most accurate implementation, please check the original repository.

Model overview

The code uses Cascaded Pyramid Network (CPN), which wins the 2017 COCO Keypoints Challenge. In Shiyu's implementation, he made a variety of modification and found two models perform the best: 1) CPN with pretrained ResNet-152 backbone and 2) CPN with pretrained SENet-154.

Here are some additional literatures that I've found useful to understand the models:

• Feature Pyramid Networks for Object Detection
• Training Region-based Object Detectors with Online Hard Example Mining

Requirements

Here are the key libraries that I've used to run the models:

• Python 3.6.5
• CUDA Toolkit 9.0+
• torch 0.4.0
• cv2
• pandas
• numpy
• fire
• nvidia-ml-py3
• py3nvml
• visdom

The easiest way to run the code is to use Docker image provided by FloydHub: docker pull floydhub/pytorch:0.4.0-gpu.cuda9cudnn7-py3.30. You'll need nvidia-docker if you're running Docker on GPUs.

Data Preparation

I follow the same directory structure as gathierry/FashionAI-KeyPointsDetectionOfApparel/. You can download FashionAI dataset from here (Login required).

Make sure to change the proj_path and db_path in utils/config.py.

Training / Evaluation / Prediction

Below are sample commands for running the models. Feel free to change/add keyword arguments by looking at utils/config.py.

To train:

python3 trainval.py main --category=skirt --model=cpn-senet --lr=1e-5
python3 trainval.py main --category=outwear --model=cpn-resnet
python3 trainval.py main --category=blouse --model=ensemble --batch_size=8

To evaluate:

python3 evaluate/predict_one.py main --category=outwear --model=cpn-resnet
python3 evaluate/predict_ensemble.py main --category=dress --model=ensemble

To predict (for submission):

python3 submission/predict_one.py main --category=outwear --model=cpn-senet
python3 submission/predict_ensemble.py main --category=dress --model=ensemble