IIAE(Interaction Information AutoEncoder) - Pytorch Implementation

Pytorch Implementation of Variational Interaction Information Maximization for Cross-domain Disentanglement(IIAE)

This implementation is focused on Image-to-Image translation task of IIAE.

Official Repository(tensorflow implementation) is here.

---

Dependencies

To resolve the dependencies manually, install

• python >= 3.6
• torch==1.7.1+cu110
• torchvision==0.8.2+cu110
• json
• tensorboard

Or if you are using Anaconda, simply run command

# create same conda environment

conda env create -f requirements.txt

---

Datasets

Datasets

1. MNIST-CDCB
2. Cars
3. AFHQ (Stil working for high quality result)

Download datasets

for MNIST-CDCB,

# Download MNIST-CDCB dataset
sudo bash download_mnist_cbcd.sh

for Cars,

# Download Cars dataset
sudo bash download_mnist_Cars.sh

for AFHQ, (download.sh file is from stargan v2 official repository)

# Download AFHQ dataset
sudo bash download.sh afhq-dataset

Preprocess datasets

MNIST-CDCB dataset and Cars dataset need to be preprocessed. You can preprocess the datasets for the network by running the codes in 'dataloader' directory

# for MNIST-CDCB,
python preprocess_mnist.py

# for AFHQ,
python preprocess_cars.py

When you try to run the preprocessing code, permission error can be occured. The error can be resolved by this command.

sudo chmod 0777 <directory>

---

Train / Test Network

With following command and config file, you can train and test the network.

python main.py <config file path>

Config

Here is the sample config file. Please follow the template.
Set train: 1 for train, train: 0 for test.

{
    "id": "IIAE_MNIST",
    "model": "IIAE",
    "dataset": "MNIST-CDCB",
    "path_to_data": "./MNIST-CDCB",
    "train" : 1,  
    "resolution": 256,
    "training": {
        "max_iters": 50000,
        "resume_iters": 0,
        "restored_model_path": "",
        "batch_size": 16,
        "lr": 0.00002,
        "weight_decay": 0.0,
        "beta1" : 0.9,
        "beta2" : 0.999,
        "milestones": [10000, 30000], 
        "scheduler_gamma": 0.5,
        "rec_weight": 100.02,
        "X_kld_weight": 5.02,
        "S_kld_weight": 0.02,
        "inter_kld_weight": 5,
        "print_freq": 100,
        "sample_freq": 2000,
        "model_save_freq": 5000
    },
    "test":{
        "test_iters": 50000,
        "batch_size" : 16,
        "test_path": "./2021-08-06_22-14_IIAE_MNIST"
    }
}

---

Sample Results

1. MNIST-CDCB

Domain Specific representation : Color Foreground / Background
Domain Shared representation : Digit Identity

• X: Colored background
• Y: Colored foreground(digit)

Experiment1 : < X to Y with different z_y(foreground color) >

X(input)	Outputs with diffferentt z_y	X to Y(reference)	Y(reference)

Experiment2 : < Y to X with different z_x(background color) >

Y(input)	Outputs with diffferentt z_x	Y to X(reference)	X(reference)

Experiment3 : < X to Y with different digits >

Digit Identity from X, Foreground Color from Y

X(Digit Identity)	Output	Y(Foreground Color)

Experiment4 : < Y to X with different digit identity >

Digit Identity from Y, Background Color from X

Y(Digit Identity)	Output	X(Background Color)

2. Cars

Domain Specific representation : Rotated Angle
Domain Shared representation : Car Identity

• X: Front view
• Y: View in 23 different degrees

Experiment1 : < X to Y with different z_y(Angles) >

X(input)	Outputs with diffferentt z_y	X to Y(reference)	Y(reference)

Experiment2 : < Y to X with different z_x(front) >

Y(input)	Outputs with diffferentt z_x	Y to X(reference)	X(reference)

Experiment3 : < X to Y with different car identity >

Car Identity from X, Rotated Angle from Y

X(Car Identity)	Output	Y(Rotated Angle)

Experiment4 : < Y to X with different car identity >

Car Identity from Y, Rotated Angle from X

Y(Car Identity)	Output	X(Rotated Angle)

3. AFHQ

• Leftmost column: Input Image
• Rightmost column: Reference image

Results with 1000000 iterations

Cat to Dog

Dog to Cat

---

References

1. https://github.com/gr8joo/IIAE
2. https://github.com/clovaai/stargan-v2
3. https://github.com/SKTBrain/DiscoGAN
4. https://github.com/agonzgarc/cross-domain-disen