A tensorflow implementation of a Learned Representation of artistic style
Most of the code in the repo was taken from hwalsuklee/tensorflow-fast-style-transfer.
This is the implementation used in my app Spectrum - Artistic Photo editor. If this fork helps you, please consider downloading the app, it really helps me out.
- Added Conditional Instance Normalization which enables a single model to learn multiple styles with only a few extra parameters.
- Used Upsampling+Convolution instead of Transpose Convolutions(prevents checkerboard artifacts).
- Experimented with Depthwise separable convolutions instead of regular convolutions. The current model still uses all convolutional layers but you can replace
_conv_layerwith_depthwise_conv_layerfor some layers to reduce the model size(but don't do this for the initial layers).
I trained my models in this Kaggle kernel. Please see run_train.py and run_test.py for training options.