Keras 2 + Tensorflow implemenation of fast-style-transfer
keras 2
Tensorflow
keras_contrib (for InstanceNormalization)
Download and extract http://msvocds.blob.core.windows.net/coco2014/train2014.zip (~13GB), into a directory called data so that the images are in 'data/train2014'. Because of the way Keras' ImageDataGenerator looks for images, the images need to be in a subdirectory of the directory passed as "train-path".
python train.py --style wave.jpg --model-output wave.h5
--style STYLE style image path
--model-output MODEL_OUTPUT path to save the trained model out as a h5 file
--model-input MODEL_INPUT path to model to train (if continuing training)
--test TEST test image path, if given will style this image
after every test-increment and save into test-dir
--test-dir TEST_DIR test image save dir
--test-increment TEST_INCREMENT number of batches to test after
--train-path TRAIN_PATH path to training images folder (default 'data')
--epochs EPOCHS num epochs (default 2)
--batch-size BATCH_SIZE batch size (default 4)
--steps-per-epoch BATCH_SIZE number of batches of samples per epoch,
should be # of samples / batch size
--content-weight CONTENT_WEIGHT content weight (default 15.0)
--style-weight STYLE_WEIGHT style weight (default 100.0)
--tv-weight TV_WEIGHT total variation regularization weight (default 200.0)
python evaluate.py --model wave.h5 --input doge.jpg --output doge-wave.jpg
--model MODEL model path
--input INPUT input image path
--output OUTPUT output image path
-p, --pad add reflection padding to input image
-b. --border-size BORDER_SIZE border size of reflection padding