Those code implement the paper Semantic Image Inpainting with Deep Generative Models First of all, we train the dcgan and finanly we train better z. We use z fill into the pre-train dcgan model and generate the uncompletion part of images.
At first, git clone my code. We use tensorflow 1.4 and python3.5. Inorder to do images preprocess, we need to use a python library: face_recognition to locate the face of image.
pip3 install face_recognition
And all the config we can change in file config.py.
We use CelebA Dataset, then we need to download img_align_celeba.zip and unzip all the images to the file img_align_celeba. Choose 300 pictures randomly as testing pictures and move them to file pre_uncompletion_image. Because all the picture is original people pictures and our model need the face pictures whose shape is 64 * 64 * 3, so that we need to use face_recognition library to locate where is the face and cut them out as same height and wide, finally, we scale them to 64 * 64 size. We need to need to run image_pro.py in the file img_align_celeba to generate images which is for training in the file image and run uncompletion_image_pro.py in the file pre_uncompletion_image in the file uncompletion_image Now, all the pictures in the image and in pre_uncompletion_image are 64 * 64 * 3.
cd img_align_celeba
python3 image_pro.py
cd ..
cd pre_uncompletion_image
python3 uncompletion_image_pro.py
cd ..
All the DCGAN details are in file model.py including dcgan model and completion function. To train dcgan, we need to run:
python3 train_dcgan.py
We will train 20 epoch in DCGAN and this is enough for our purpose.
After trained DCGAN, we need to train z and for my code, I will randomly choose 64 image in the file pre_uncompletion_image and run for testing.
python3 train_completion.py
All the outcome is in file completions which we will generate in our code.
d_loss_fake
d_loss_real
d_loss
g_loss
the example
the example with mask
after completion
