Developer shares LSTM training struggles in C++

A developer discusses challenges encountered while implementing a 2-layer LSTM in C++, noting that the model often fails to converge and exhibits erratic error coefficients during training.

The user reports using a 56k character dataset with 96-node input layers, observing that the error rate frequently climbs past the 3.0 threshold despite attempts at gradient normalization and clipping.

Although the architecture successfully overfits on elementary tests, the author seeks advice on improving stability without restructuring the code to a single hidden layer.