Interactive Fashion Recommendation Dataset

A new interactive dataset for fashion recommendation tasks.

Introduction

This dataset is build upon Polyvore dataset. Concerning of the possible copyright issue, we provide a script to generate the dataset with modifying the original data.

Preparation

1. Python 3.5+ with pandas, tqdm
2. Polyvore dataset [original repo]
   • dataset ~ 8 MB
   • images ~ 13GB
   • Gather the above two parts into a directory:

     polyvore
     ├─images
     │  ├─123456(outfit id)
     │  └─...
     ├─test_no_dup.json
     ├─train_no_dup.json
     └─valid_no_dup.json
     
     

Usage

Use python scripts/generate.py -h to check the help message anytime.

usage: generate.py [-h] --polyvore POLYVORE [--record RECORD] [--target TARGET] [--move MOVE]
                   [--override_category OVERRIDE_CATEGORY]

optional arguments:
  -h, --help            show this help message and exit
  --polyvore POLYVORE   The path of Polyvore dataset
  --record RECORD       The path of interactive dataset id record
  --target TARGET       The path where the generated dataset saves
  --move MOVE           Use "move" rather than "copy" operation to the images in Polyvore dataset
  --override_category OVERRIDE_CATEGORY
                        Set true to sort the categories into 7 main groups

We also provide example file in generate.sh.example(for Linux & macOS) and generate.bat.example (for Windows NT). Before running the shell / batch file, do not forget to remove the .example extension from the file name and add your python environment commands.

Citation

Feel free to cite our paper of this dataset:

bibtex code