onmt-runner

onmt-runner loads legacy (OpenNMT-py 1.x-style) recurrent .pt checkpoints without installing OpenNMT-py or its obsolete torchtext dependency.

from onmt_runner import Translator

translator = Translator.load("murre_norm_default.pt")
print(translator.translate("h e i"))
print(translator.translate_many(["h e i", "m i t ä"]))

OpenNMT-tf SavedModel directories use the same loading entry point after installing the optional TensorFlow dependency:

python -m pip install 'onmt-runner[tensorflow]'

from onmt_runner import Translator

translator = Translator.load("flags_dist")
print(translator.translate("Kainuu h e i"))

Some older OpenNMT-tf exports use the retired TensorFlow Addons GatherTree operation. onmt_runner transparently compiles a small CPU compatibility op on first use; a C++17 compiler such as g++ must be available for those exports.

Input must already be tokenized exactly as it was during training. For a character model, that normally means characters separated by spaces. Returned tokens are joined with spaces by default; pass detokenize=lambda tokens: "".join(tokens) to translate or translate_many for character output.

Supported checkpoints

The initial release supports the common OpenNMT-py 1.x text checkpoint format whose opt.decoder_type is rnn, including bidirectional LSTM encoders, input-feeding decoders, and dot/general global attention. It deliberately does not execute serialized OpenNMT code: only tensor weights and vocabulary metadata are read. As with torch.load(..., weights_only=False), only load checkpoints you trust: historic OpenNMT checkpoints use Python pickle.

Transformer, copy-attention, and non-text checkpoints raise a precise UnsupportedCheckpointError rather than producing unreliable output.

Building a wheel

python -m pip install -r requirements.txt
python setup.py bdist_wheel

The resulting wheel is written to dist/. A standards-based alternative is python -m pip install build && python -m build.