Improved INT8 quantization to be performed based on saved_model signature information Ref: https://www.tensorflow.org/lite/performance/post_training_quantization#full_integer_quantization Improved matching of the input OP name specified in the dataset for calibration with the input OP name in the signature to prevent input order discrepancies. When quantizing INT8 for models with multiple inputs, you no longer need to be aware of the order in which the calibration data sets are specified. Only when performing INT8 quantization, the saved_model signature information of the converted model is displayed in the log as reference information, as shown in the figure below. The input OP name in ONNX and the input OP name after conversion to saved_model may mismatch. This is due to automatic sanitization of strings that cannot be used in the input OP name of saved_model. e.g. :, / [BERT-Squad] INT8 quantization: The input data type must be Float32. #248 What's Changed Improved INT8 quantization to be performed based on saved_model signature information by @PINTO0309 in https://github.com/PINTO0309/onnx2tf/pull/249 Full Changelog: https://github.com/PINTO0309/onnx2tf/compare/1.7.26...1.7.27

If you use onnx2tf in your research, please cite it using these metadata.