cutecave/bopodoc-models

G2PW int8 (ONNX)

g2pw_int8.onnx is an int8-quantized copy of the G2PW model for Mandarin polyphone disambiguation (picking the right reading of a character from its context). It is used by BopoDoc with ONNX Runtime Web.

Source

• G2PW: g2pW: A Conditional Weighted Softmax BERT for Polyphone Disambiguation in Mandarin, Yi-Chang Chen et al., INTERSPEECH 2022. https://github.com/GitYCC/g2pW, Apache License 2.0.
• G2PW is fine-tuned from bert-base-chinese, Apache License 2.0.

Changes from the original

This file is a modified version of the FP32 ONNX model (g2pw.onnx) distributed with g2pW:

• Weights were quantized from FP32 to int8 with ONNX Runtime dynamic quantization (quantize_dynamic, QuantType.QInt8). Activations stay FP32.
• No retraining. Inputs, outputs, labels and vocabulary are unchanged.

License

Apache License 2.0. See LICENSE.