Positional Encoding

In NeRF they showed that their neural network model perform poorly at reconstructing high frequency details when directly operating on coordinates. They were able to achieve significantly higher performance when encoding coordinates as a vector of sinusoids. torchsample.encoding contains some common coordinate encoding methods.

All encoding functions take in a single required argument, coords and produce a new transformed coordinate tensor.

coords = ts.coord.rand(16, 4096, 2)
gamma_encoded_coords = ts.encoding.gamma(coords)  # (16, 4096, 40)