Dataset

The dataset for this repo can be downloaded here.

Place the downloaded file in the folder data/

Training and evaluating

Training (need to be updated later)

python main.py

positional arguments:

patch_size            Patch size for D

optional arguments:

-h, --help            show this help message and exit
--backend BACKEND     theano or tensorflow
--generator GENERATOR
                    upsampling or deconv
--dset DSET           facades
--batch_size BATCH_SIZE
                    Batch size
--n_batch_per_epoch N_BATCH_PER_EPOCH
                    Number of training epochs
--nb_epoch NB_EPOCH   Number of batches per epoch
--epoch EPOCH         Epoch at which weights were saved for evaluation
--nb_classes NB_CLASSES
                    Number of classes
--do_plot             Debugging plot
--bn_mode BN_MODE     Batch norm mode
--img_dim IMG_DIM     Image width == height
--use_mbd             Whether to use minibatch discrimination
--use_label_smoothing
                    Whether to smooth the positive labels when training D
--label_flipping LABEL_FLIPPING
                    Probability (0 to 1.) to flip the labels when training
                    D

Example:

python main.py 64 64

Output

• The weights are stored in models/
• The generated images are stored in figures/
• Save model weights every few epochs

Additional notes

You can choose the type of generator:

• The image dimension must be a multiple of the patch size (e.g. 256 is a multiple of 64)
• In the discriminator, each patch goes through the same feature extractor. Then the outputs are combined with a new dense layer + softmax
• upsampling: generate the image with a series of Upsampling2D and Convolution2D operations
• deconv: use keras' transposed convolutions Deconvolution2D. This is closer to the original DCGAN implementation.