Changing the learning rate

[ooocct]

Hi, Denis. Thanks for your effort on this lib. I am trying to use this in my research. However, it seems the learning rates of GRU and LSTM are all fixed after their initialization. I want to change them after every epoch. Could you please give me some advice? Regards. Gin

[steckdenis]

Hello,

The LSTM and GRU units are built from a set of operations, with every learnable weights in Dense layers (so, there are a few Dense layers as part of an LSTM or GRU unit). In Dense, the learning rate is indeed constant after construction, but you can try to implement a simple learning rate schedule by modifying _learning_rate in Dense::update.

There is unfortunately no way to change the learning rate from outside of nnetcpp, as this is not a feature I envisioned for this (quite minimal) neural network library. By the way, if you need speed, you can have a look at the PyTorch C++ API, that allows you to build neural networks, forward them and autograd them directly from C++. PyTorch implements neural networks based on optimized matrix multiplication routines, while nnetcpp (its Dense layer) uses a matrix-vector product that is less efficient.