├── README.md
└── gan-model-on-anime-image-dataset-tensorflow.ipynb


/README.md:
--------------------------------------------------------------------------------
 1 | # GAN-model-using-TensorFlow
 2 | In this repo, I will use GAN(Generative Adversarial Network) on Anime_image dataset.
 3 | GAN is short form of Generative Adversarial Network and a deep learning architecture. GAN consists of 2 parts, Discriminator and Generator.
 4 | The Generator tries to creat fake images and fool the Discriminator, and Discriminator tries to distinguish the images and label them as fake(0) or real(1).
 5 | 
 6 | This zero-sum game continuees until the Generator can no longer creat images which fools the Discriminator and the Discriminator cannot be fooled.
 7 | 
 8 | There are different types of GAN Models but we are using DCGAN which is the short form of Deep Convolutional GAN.
 9 | 
10 | 
11 | 
12 | **What is Discriminator ?**
13 | 
14 | The Discriminator is a Neural Network model which tries to distinguish the real images from fake images(generated by Generator) and label them as fake(0) or real(1).
15 | 
16 | Notes :
17 | 
18 | 1.The image size is (64,64), so the input_shape of first conv2d layer should be (64,64,3).
19 | 
20 | 2.The output of Discriminator is either a 0(fake) or 1(real).
21 | 
22 | 3.Using "same" as padding ensures us that the output dimension is not going to change.
23 | 
24 | 4.In the Discriminator function, all activations should be "LeakyReLU", exept the last layer which should be "sigmoid"
25 | 
26 | 5.The last layer is using "sigmoid" as activation function to create a binary output, which real images are labeled as 1 and the fake ones are labeled as 0.
27 | 
28 | 6.The Discriminator downsamples the input shape.
29 | 
30 | 
31 | **What is Generator ?**
32 | 
33 | The Generator tries to creat fake images and fool the Discriminator, and Discriminator tries to distinguish the images and label them as fake(0) or real(1).
34 | 
35 | Notes :
36 | 
37 | 1.The latent space is an arbitrarily defined vector space of Gaussian-distributed values and here I consider 100 as latent_dim.
38 | 
39 | 2.Also units in Dense layer can be (4,4,256) (so it has 4096 nodes)---> 256 versions of 4*4 images.
40 | 
41 | 3.Using "same" as padding ensures us that the output dimension is not going to change.
42 | 
43 | 4.The output shape will be (None, 64, 64, 3) just like the input image(real image) of Discriminator.
44 | 
45 | 5.In Generator, we can use Upsampling or Conv2DTranspose layer to upsample the input.
46 | 
47 | **This dataset has 63,632 "high-quality" anime faces.**
48 | The source of the dataset is from kaggle : https://www.kaggle.com/datasets/splcher/animefacedataset
49 | 


--------------------------------------------------------------------------------