Set neural network initial weight values in C++ torch

Question

I was looking for an API to set initial weight values in libtorch. In the python version, (i.e. pytorch) one can easily use torch.nn.functional.weight.data.fill_(xx) and torch.nn.functional.bias.data.fill_(xx). But, it seems that such an API does not exist in C++ yet. I would appreciate any help or comment to achieve such functionality.

Thanks, Afshin

Answer 1

I got this solution better than the previous one in which model is an object of type torch::nn::Sequential:

torch::NoGradGuard no_grad;

for (auto &p : model->named_parameters()) {
    std::string y = p.key();
    auto z = p.value(); // note that z is a Tensor, same as &p : layers->parameters

    if (y.compare(2, 6, "weight") == 0)
        z.uniform_(l, u);
    else if (y.compare(2, 4, "bias") == 0)
        z.uniform_(l, u);
}

Instead of uniform_ you can use normal_, ... which are available on torch. This solution is not limited to torch::nn::Linear layer and can be used foy any layer type.