Collect infos of features by running model in onnxruntime :param model: the model :param image_dir: the directory of images :param fe

since flatbuffers doesn't have float16 data type, use uint16 instead

There are some requirement for quantization info, we assert and infer them here Some layer sequence need multiple runs to make infos right, l

Modify proto buffers when all quantization infos are set correctly :param m: the model :param quant_layers: layers need to be quantized

values of initializers may be stored in float_data or raw_data, if in raw_data, we move them to float_data for convenience :param m: the

quant weights before collecting min and max, for simulating the effect of quantization :param m: the model :param quant_layers: layers ne

a float32 array of scales, the length will be 1 for non per-channel quantization, and be channelDim for per-channel quantization

NNAPI requires scales[bias] equals scales[input]*scales[weight] and zps[scale]=0 :param m: the model :param quant_layers: layers need to

Add /dev/shm implementation of shared memory for non-Android platforms

TODO: combine it and AddInitializersFromBuffer

Methods for backward compatibility

ModelBuilder auto generated methods start

* @brief Add NNAPI operand from `buffer`, the memory pointed * by `buffer` should be persistent until the execution finished. * No copying. * * @p