Another type of GAN, using Deep Convolutional Generative Adversarial Nerworks.
-
The structure is based on Deep Convolutional Generative Adversarial Nerworks, worked on TensorFlow.
-
The whole training takes much time and calculation amount, maybe you'd better use GPU version or you will have to wait for a long time training.
-
Another choice is that you can download the trained model by me. I will upload the model trained from Celeb's 10000 face pictures.
-
In
DCGAN_64x64.py
, used as the structure:while in
DCGAN_32x32.py
, I reduce one layer so the output is 32x32x3 pic. -
Here is the hyperparameter:
Then there I recorded the training process in 4 kind of training dataset:
The results using fully connected network is:
while using this DCGAN is :
Here is the results in different training dataset:
- On drawing figure face:
- On car pictures:
- On human face pictures:
- On coins:
there is much difference, right?
-
Put all datasets in to a folder and change the
input_data
content to your dataset folder, After it trained, you can turn off To_Train button and run it again it will create pictures generated by the generator. -
I will uploaded trained model to the folder
训练完成的模型
, whose name represent the dataset's type I trained. Also, I will put the hyperparameter txt in it so that you should keep yours same to that and you can use it.
I strongly recommend to use 64x64 code for I have updated it in continuing training and learning rate decay, and the model I upload will be that code.