Personalized Fashion Recommendation and Generation

This is our TensorFlow implementation for the paper:

Wang-Cheng Kang, Chen Fang, Zhaowen Wang, Julian McAuley. Visually-Aware Fashion Recommendation and Design with Generative Image Models. In Proceedings of IEEE International Conference on Data Mining (ICDM'17)

Please cite our paper if you use the code or datasets.

We provide the three modules in our framework:

Deep Visually-Aware Bayesian Personalized Ranking (DVBPR): Jointly learn user latent factors and extract task-guided visual features from implicit feedback for fashion recommendation.
GANs: A conditional generative adversarial network for fashion generation.
Preference Maximization: Adjust generated images that match a user's personal taste better (personalized fashion design).

Environment

The code is tested under a Linux desktop with a single GTX-1080 Ti GPU.

Requirements:

TensorFlow 1.3
Numpy
PIL

Datasets

The four fashion datasets:

AmazonFashion (3.3GB) : 64K users, 234K images, 0.5M actions
AmazonWomen (6.2GB): 97K users, 347K images, 0.8M actions
AmazonMen (2.1GB): 34K users, 110K images, 0.2M actions
Tradesy (3.4GB): 33K users, 326K images, 0.6M actions

can be downloaded via

bash download_dataset.sh

All datasets are stored in .npy format, each item is associated with a JPG image. Please refer to DVBPR code for detail usage. For image generation, we mainly use the AmazonFashion dataset.

Amazon datasets are derived from here, tradesy dataset is introduced in here. Please cite the corresponding papers if you use the datasets.

Please note the raw images are for academic use only.

Model Training

Step 1: Train DVBPR:

cd DVBPR
python main.py

The default hyper-parameters are defined in main.py, you can change them accordingly. AUC (on validation and test set) is recorded in DVBPR.log.

Step 2: Train GANs:

cd GAN
python main.py --train True

The default hyper-parameters are defined in main.py, you can change them accordingly. Without '--train True', it will load a trained model and generated images for each category (stroed in folder samples).

Step 3: Preference Maximization:

cd PM
python main.py

PM is based on pretrained DVBPR and GAN models. It will randomly pick a user for each category, and show the generated images through the optimization process.

With a single GTX-1080 Ti, training DVBPR and GANs take around 7 hours respectively.

Demo (with pretrained models)

A quick way to use our model is using pretrained models which can be acquired via:

bash download_pretrained_models.sh

With pretrained models, you can see the AUC results of DVBPR, and run GAN and PM code to generate images.

Misc

Acknowledgments: GAN code borrows heavily from DCGAN. GAN networks are modified from LSGAN.
In principal, our framework can adapt with any GANs variant, we look forward to using advanced GANs to achieve better generation results with higher resolution.

kang205 / DVBPR