how can I freeze subset of layers of NN from training in C++? OR is there a way to only pass subset (some parameter group)of model params to the optimizer for gradient updates in C++?
Thanks in advance.
how can I freeze subset of layers of NN from training in C++? OR is there a way to only pass subset (some parameter group)of model params to the optimizer for gradient updates in C++?
Thanks in advance.
You could be able to iterate the parameters and set their requires_grad
attribute to False
.
Something like this would work where module
is the desired layer:
for (auto& parameter : module->parameters()) {
parameter.requires_grad_(false);
}