README.md
We use a neural network that takes a coordinate (x, y) and output the corresponding RGB value to generate an image.
With different random hyperparameters, the network generates images of large style variance.
The image is generated using a single-path feed-forward network with only fully-connected layers and batch-normalization, flavored with various nonlinearities and initialized with gaussian random weights.
The key to generate these intriguing images is the use of random nonlinearities from a large nonlinearity pool at each layer, especially those discontinuous functions (like sign) or functions with large gradient near some point (e.g, log(abs(x) + eps)).
Here are the nonlinearities we've used so far:
NONLIN_TABLE=dict(relu=T.nnet.relu,tanh=T.tanh,abs_tanh=lambdax: abs(T.tanh(x)),sigmoid=T.nnet.sigmoid,softplus=T.nnet.softplus,sin=T.sin,cos=T.cos,sgn=T.sgn,sort=lambdax: T.sort(x, axis=1),abs=abs,log_abs=lambdax: T.log(abs(x) + eps), # this is awesomelog_abs_p1=lambdax: T.log(abs(x) +1),log_relu=lambdax: T.log(T.nnet.relu(x) + eps),log_square=lambdax: T.log(x**2+ eps), # just a scalarxlogx_abs=lambdax: T.xlogx.xlogx(abs(x) + eps),xlogx_abs_p1=lambdax: T.xlogx.xlogx(abs(x) +1),xlogx_relu=lambdax: T.xlogx.xlogx(T.nnet.relu(x) + eps),xlogx_relu_p1=lambdax: T.xlogx.xlogx(T.nnet.relu(x) +1),xlogx_square=lambdax: T.xlogx.xlogx(x**2+ eps),softmax=T.nnet.softmax,logsoftmax=T.nnet.logsoftmax,hard_sigmoid=T.nnet.hard_sigmoid,identity=lambdax: x,square=lambdax: x**2
)See https://github.com/zxytim/neural-painter/blob/master/neural_painter.py for implementation. The code is based on theano.
./neural_painter.py --image_size 800x800 --hidden_size 100 --nr_hidden 4 --nonlin random_every_time --nr_channel 3 --output_nonlin identity --coord_bias --seed 42 --output 42.png- Wikipedia page for Compositional pattern-producing network
- CPPN in Tensorflow
- Image regression from karpathy:
- Generating Abstract Patterns with TensorFlow:
- High resolution MNIST generator (CPPN + GAN/VAE)