aidsoid/punct_cap_seg_47_language-CoreML-fp32

punct_cap_seg_47_language, Core ML, fp32

A Core ML conversion of 1-800-BAD-CODE/punct_cap_seg_47_language: punctuation, true-casing and sentence boundary detection for 47 languages, in one pass over lower-cased, unpunctuated text.

Converted from the original ONNX model at commit 1b9d51fc7989ebc61e844d407d9dadd08ff4ba28 (ONNX → PyTorch with onnx2torch → Core ML with coremltools 8.3) by conversion/convert.py, with the versions in conversion/requirements.txt. The weights stay fp32: half precision changes the predictions. The script checks that the punctuation and sentence ends it predicts match the ONNX model's.

Files

• PunctCapSeg47.mlpackage: the model. Input input_ids, int32 [1, 128]: <s> (1), the text's token ids, </s> (2), padded with <pad> (3). Outputs, one per token:
  • pre_preds, post_preds: indices into pre_labels and post_labels of the original config.yaml — the mark before and after the token;
  • cap_preds ([1, 128, 16]): 1 where that character of the token is upper case;
  • seg_preds: 1 where a sentence ends after the token.
• vocab.json: the SentencePiece unigram vocabulary (piece, score, type) of spe_unigram_64k_lowercase_47lang.model, and the special ids.
• ORIGINAL_README.md: the original model card.
• conversion/convert.py, conversion/requirements.txt: the conversion.

License

Apache 2.0, as the original model. All credit for the model goes to its author.