Tensorflow 2.0 implementation of fourier feature mapping networks.

Based on the paper "Fourier Features Let Networks Learn High Frequency Functions in Low Dimensional Domains".

Results after training fourier feature mapping models for around 100-200 epochs.

• Git clone
• Install the package locally (the -e flag means that any changes you make here in the code will be reflected in the package)

  pip install -e .

Check out the 2D image regression example for an example usage.

Run the example using:

python ./fourier_dnn/2d_image_regression.py 3 100

where 3 is the image index from the dataset (of 16 images) and 100 is the number of epochs. (Dataset taken from the original implementation)

Use fourier_dnn.ffm_mlp.FourierMLP to make a complete fourier feature mapping network. For example:

model = FourierMLP(10, 256, 3,
                   gaussian=True, staddev=16,
                   num_units_FFM=256)

Use fourier_dnn.metrics.PSNR as a metric for model.compile to get PSNR values on every epoch.

• Effect of standard deviation
• Effect of learning rate
• Effect of number of Dense units
• Effect of standard deviation 2
• Effect of number of layers