This is one of Deep learning nanodegree major projects. In this project, I have trained generative adversarial networks, especially DCGAN to generate new images of faces.My goal was to get a generator network to generate new images of faces that look as realistic as possible!
I used CelebFaces Attributes Dataset (CelebA) to train your adversarial networks.
-
I have tried varuios moedel architicture :
1- For model depth : [3 , 4] layers
2- For convolution layer deimention : [32 , 64]
-
I have tried also number of batch size : [32 , 64 , 128 ,256]
-
I also have tried diffrient number of epochs : [ 20 , 30 , 50]
-
As in the practise notebooks i used adam optimizer ,as it was recommended.
Actually , the models has smal deffriences and can't appear from the few sample of photos i show , but we can deffirenciate between them by the loss perfornmence.As we can see in the above graphs , the best model with loss performence is the model with the following architicutre :
layer_in_depth = 4
conv_dim = 64
batch_size = 64
num_epochs = 10
To hava a closer look to the results of each model and its traing loss plot, i encourge you to have a look at the Training loss and generator sample of training in the following report