Disco-GAN-Tensorflow
This repositroy shows fashion items data orientated Disco-GAN implementation by using tensorflow.
-Original Research Paper: Leraning to Discover Cross-Domain Relations with Generative Adversarial Networks
-Related github Repo: SKTBrain/Disco-GAN
Paper-summary
- Labeling and paring datas are costful and labor intensive
- By given 'unpaired data' GAN finds relations btw two diff domains
- No pre-trained model required
- Two diff GAN coupled together
Create own dataset
- Construct feasible dataset requires intensive efforts. Following few ideas help you to build dataset
- Offical paper's data uses at least 50,000 images(well organized and well formed) per item.
- Use authentic and reliable crawler: recommend to use AutoCrwaler
- In keyword.txt, lists up auto-generated tags(from google image search) with original item that you are looking for
- EX) phone case, phone case aztec, phone case pattern, phone case flower, etc
- This may help to build your dataset more robust and enough to be taken by trainning model
For more detail infos such as prerequisites, code descriptions, params setting, db setting, followed this link.
1. Train
python3 train.py --train_A <directory-first-database> --train_B <directory-second-databse --epochs <#> --batch_size <#>2. Training Result
- First trial: using edges2handbags(first) and edges2shoes(second) - around 49,000 imgs(SHOES) - 130,000 imgs(HBG) - 30 epoch
- Second trial: using clutch bag(first) and sandals(second) - around 1,300 - 1,400 images per items - 200 epochs - 19 steps
-
Third trial: using brand logo(first) and soccer shoes(second) - around 5,700 images(shoes) - 8,000(brand logo)
-
4th : (upcoming) using backpack and smartphone case (Collecting dataset at least 50,000+ images)
