Hyukjin Jeong

In https://github.com/Samsung/ONE/pull/13585, `CircleQuantParam` is not updated.

#12063 is merged, so this issue can be closed?

I used the below script to compare the result of onnx and tflite. ``` $ onecc import onnx -i bn.onnx -o bn.circle --save_intermediate ``` ```python import tensorflow as tf import...

Constant values are changed when onnx is converted to tflite. In onnx, they are mean/variance, but in tflite, they are channelwise mul/add. > I'm curious if onnx-runtime does something additional...

Yes. Thank you :)

Here is a useful code snippet. ``` model = tf.saved_model.load(saved_model_dir) concrete_func = model.signatures[ tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY] concrete_func.inputs[0].set_shape([1, 512, 2049, 2]) converter = TFLiteConverter.from_concrete_functions([concrete_func]) converter.allow_custom_ops = True converter.experimental_new_converter = True converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS]...

We now have `dynamic_batch_to_single_batch` option in one-optimize. Let's close this.

Looks good to me. > IR: go with CircleCirGru It would be nice to add comments why "Cir" is repeated near the `CircleCirGru` class definition. > circle2circle option fuse_gru looks...

## Information about custom Ops ### OnnxQuantizeLinear, OnnxDequantizeLinear - Inputs - Input - Scale - Zeropoint - attributes - T: output dtype (based on circle schema) - axis: quantized_dimension We...

> Propagate qparam: There are some missing quantization points. It is necessary to propagate qparams for some Ops. -> Need to be done in one-quantize List of operators to quantize...