lvapeab/nmt-keras

Implements both attention-based RNN and Transformer architectures with multi-GPU training, beam search decoding, and ensemble inference with length/coverage normalization. Supports multiple attention mechanisms (Bahdanau, Luong, double stochastic), conditional GRU/LSTM units, interactive translation protocols, and integration with pretrained embeddings (GloVe, Word2Vec). Built on Keras/TensorFlow with client-server architecture for web deployment and Tensorboard monitoring.

531 stars. No commits in the last 6 months.