Gender Changer w/ CycleGAN 👫

👤 Author : Myoons

• Github
• Website

🌈 Motivation : To give pleasure to friends

FUN is the best! It's always thrilling.

📝 Dataset : Korean Celebrities

Since the purpose of the project was to give pleasure to my friends, I needed a model that was appropriate to Koreans. So I collected Korean Celebrities Images as Dataset. I crawled images with Selenium and used Face Recognition to crop the face in the images.

Man : 3667 Images

Woman : 4272 Images

Dataset Link

🔧 Training

1. Install Repository

You can install this repository using git clone

git clone https://github.com/myoons/CycleGAN-Gender-Changer.git

2. Setup Dataset

Download the dataset from Google Drive
Then build the dataset by setting up the following the directory structure.

├── datasets
|   ├── <dataset_name>         # i.e. genderchange
|   |   ├── train              # Training
|   |   |   ├── A              # Contains domain A images (i.e. Man)
|   |   |   └── B              # Contains domain B images (i.e. Woman)
|   |   └── test               # Testing
|   |   |   ├── A              # Contains domain A images (i.e. Man)
|   |   |   └── B              # Contains domain B images (i.e. Woman)

3. Train!

python train.py --dataroot datasets/<dataset_name> --cuda

This command will start a training session using the images under the dataroot/train directory with the hyperparameters that showed best results according to CycleGAN authors.

If you don't own a GPU remove the --cuda option, although I advise you to get one!

📟 Tensorboard

tensorboard --logdir ./logs

You can watch your experiments' progress by runing tensorboard

🚩 Major Flags

1. --n_epochs : Number of epochs in training

＊ Default : 400
＊ Type : Int

2. --batchSize : Size of Batch

＊ Default : 10
＊ Type : Int

3. --size : Size of Image Crop (Squre Assumed)

＊ Default : 256
＊ Type : Int

4. --dataroot : Root Directory of Dataset

＊ Default : Input as Arugment
＊ Type : Str

5. --input_nc / --output_nc : Number of channels of input / output data

＊ Default : 3
＊ Type : Int

📁 Directories

1. data_utils : Crawl & Pre-process dataset

＊ Foreign_Crawling : Crawling foreign person images
＊ Format_Change : PNG2JPG & File Renaming
＊ Korean_Crawling : Crawling Korean Celebrities images
＊ chromedriver : Selenium Chrome Driver
＊ face_detector : Croping face from images

2. models : Discriminator & Generator

＊ discriminator : Discriminator model
＊ generator : Generator Model

3. utils : Utils for training

＊ utils : training utils

📈 Learning Graph (Tensorboard)

1. Loss_D

2. Loss_G

3. Loss_G_GAN

4. Loss_G_Identity

5. Loss_G_Cycle

🔥 Results

1. Man2Woman

2. Woman2Man

🔗 References

1. Paper
2. Base Code
3. Face Recognition
4. Project Description
5. CycleGAN Home Page
6. CycleGAN Description Video