[FEATURE REQUEST] A cppn example

[skriegman]

Hi. I'd like to use this library with CPPNs that take as their input xyz and output a 3d tensor of reals.

Wondering if you have a bare bones implementation (e.g. fixed architecture) on hand that you are willing to share, and a paper you'd like to be cited if that example were to be adapted for academic research.

Thanks,
Sam

[btjanaka]

Hi Sam! Unfortunately, we don't have any CPPN examples right now, but here are some pointers:

• pyribs would only work with a fixed architecture, as we assume solutions are a fixed length
• To make a neural net work with pyribs, you would need to be able to serialize it to a 1D array, and then deserialize it back to a neural net.
• In PyTorch, this can be done by adding methods to an nn.Module like the following:

    def serialize(self):
        """Returns 1D array with all parameters in the network."""
        return np.concatenate(
            [p.data.cpu().detach().numpy().ravel() for p in self.parameters()])

    def deserialize(self, array):
        """Loads parameters from 1D array."""
        arr_idx = 0
        for param in self.model.parameters():
            shape = tuple(param.data.shape)
            length = np.product(shape)
            block = array[arr_idx:arr_idx + length]
            if len(block) != length:
                raise ValueError("Array not long enough!")
            block = np.reshape(block, shape)
            arr_idx += length
            param.data = torch.from_numpy(block).float()
        return self

• Finally, we do have this tutorial where we try to train an MNIST network with CMA-ME, but we never finished it (and it uses an outdated pyribs version)

[btjanaka]

As for citation, you can cite the pyribs library and (if you use it) CMA-ME: https://pyribs.org/citation/