Ja2_Faces_Generation

Overview

The goal of this project is to generate faces using the Generative Adversarial Network based on the original photos of the mercenaries from game Jagged Alliance 2. I used Google Colab to training models. I interupt training after 65000 epochs, when I didn't see any future progress.

Data

Images come from the game Jagged Alliance 2 version 1.13 https://jagged-alliance.pl/download72

Images Exctraction

To extract orginal mercenaries faces from this game you need:

• Unpack Faces.slf file (located in the main game data folder) using SLFExplore program http://ironsquad.pl/plik15/
• Convert all sti files to png. I writed script convert_sti_to_png.py using the https://github.com/ja2-stracciatella/ja2-open-toolset/tree/master/ja2py library to do this.
• Extract 0.png files from each mercenary folder and give them unique names. I writed script unpack images to do this.

I enclosed all above scripts in Scripts directory.

GAN Architecture

GENERATOR

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
dense_1 (Dense)              (None, 147456)            14893056  
_________________________________________________________________
reshape (Reshape)            (None, 24, 24, 256)       0         
_________________________________________________________________
up_sampling2d (UpSampling2D) (None, 48, 48, 256)       0         
_________________________________________________________________
conv2d_2 (Conv2D)            (None, 48, 48, 256)       590080    
_________________________________________________________________
batch_normalization (BatchNo (None, 48, 48, 256)       1024      
_________________________________________________________________
activation (Activation)      (None, 48, 48, 256)       0         
_________________________________________________________________
up_sampling2d_1 (UpSampling2 (None, 96, 96, 256)       0         
_________________________________________________________________
conv2d_3 (Conv2D)            (None, 96, 96, 256)       590080    
_________________________________________________________________
batch_normalization_1 (Batch (None, 96, 96, 256)       1024      
_________________________________________________________________
activation_1 (Activation)    (None, 96, 96, 256)       0         
_________________________________________________________________
conv2d_4 (Conv2D)            (None, 96, 96, 3)         6915      
_________________________________________________________________
activation_2 (Activation)    (None, 96, 96, 3)         0         
=================================================================
Total params: 16,082,179
Trainable params: 16,081,155
Non-trainable params: 1,024
_________________________________________________________________

DISCRIMINATOR

_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv2d (Conv2D)              (None, 48, 48, 128)       3584      
_________________________________________________________________
leaky_re_lu (LeakyReLU)      (None, 48, 48, 128)       0         
_________________________________________________________________
dropout (Dropout)            (None, 48, 48, 128)       0         
_________________________________________________________________
conv2d_1 (Conv2D)            (None, 48, 48, 256)       295168    
_________________________________________________________________
leaky_re_lu_1 (LeakyReLU)    (None, 48, 48, 256)       0         
_________________________________________________________________
dropout_1 (Dropout)          (None, 48, 48, 256)       0         
_________________________________________________________________
flatten (Flatten)            (None, 589824)            0         
_________________________________________________________________
dense (Dense)                (None, 1)                 589825    
=================================================================
Total params: 888,577
Trainable params: 888,577
Non-trainable params: 0
_________________________________________________________________
_________________________________________________________________

Dependencies

• Python 3.7
• Numpy
• Matplotlib
• TensorFlow 2.0 GPU Version
• Pillow
• SLF Explore (for unpacking slf files)
• Ja2py (for convert sti files)

Final generated images after 65000 epochs

The generated faces are quite realistic to me. Notice that I only used 130 good quality image to training.

Training history

As you can see after several dozen epochs, the loss of the generator and discriminator stabilized, but the quality of images continued to increase.

GAN learning progress as a GIF