Keras NMT toolkit that predates the hype cycle

What it does

NMT-Keras is a neural machine translation library built on a forked Keras 2.0.7+ and a custom multimodal wrapper. It trains and runs seq2seq models—RNN-based with Bahdanau or Luong attention, plus a Transformer implementation—through a config.py and main.py workflow. Decoding supports beam search, ensemble averaging, length normalization, and n-best list generation.

The interesting bit

The project doubles as a research vehicle for interactive and online NMT: there’s a whole branch for protocols where human translators collaborate with the model in real time, plus a client-server web demo. That’s unusual depth for a Keras-era toolkit, and the authors published on it.

Key highlights

• Supports both attentional RNNs (GRU/LSTM, conditional variants, residual stacks) and Transformer architectures
• Multi-GPU training (TensorFlow only), TensorBoard logging, and label smoothing
• Pretrained embeddings (GloVe, Word2Vec), unknown-word replacement, and model averaging
• Interactive NMT branch with online learning; live web demo at casmacat.prhlt.upv.es/inmt
• Two Colab notebooks for tutorial and model-dissection walkthroughs

Caveats

• Requires a specific fork of Keras (MarcBS/keras) and a companion multimodal_keras_wrapper—not stock dependencies
• Theano backend is untested and throws a known, non-fatal InconsistencyError
• Python 3.7 badge suggests the codebase hasn’t tracked the ecosystem forward

Verdict

Worth a look if you’re reproducing older NMT papers, studying interactive translation interfaces, or need a hackable Keras-era baseline. Skip it if you want production-scale training with modern PyTorch or JAX stacks.