JonathanRaiman/theano_lstm

Provides modular layer types (LSTM, RNN, Embedding, GatedInput) built on Theano's symbolic computation graph with support for stacked architectures and backpropagation through time via `theano.scan`. Implements gradient stabilization through element-wise clipping and Adadelta optimization to handle vanishing/exploding gradients in sequence modeling. Handles variable-length sequences through masked loss computation, enabling efficient minibatch training without sequence length uniformity constraints.

303 stars and 27 monthly downloads. No commits in the last 6 months. Available on PyPI.