fashion-compatibility

This repository contains a Pytorch fashion compatibility model.This Pytorch implementation is built on the mvasil's fashion-compatibility and rxtan2's Learning-Similarity-Conditions.There are some differences between those implementations. In particular, this Pytorch version support

• uses Resnet50 to extract the image's generalized feature.

• has a mask branch to get the cloth's shape feature.

• uses the whole outfit and multiple negative samples to calculate margin ranking loss.

Requirments

pytorch==1.2.0

torchvision==0.2.1

numpy==1.16.4

PIL==6.2.0

Usage

1. Clone this repository

   git clone https://github.com/zuoxiang95/fashion-compatibility.git
   

2. Get the train data

   There are two kinds of data.

   Polyvore data:

   Alibaba's dida data:

3. Train your model

   You can begin to train your own model with the command:

    python main.py --name 'your own training name'
   

4. Test your model

   python test_in_polyvore.py