This CL adds experimental enable_numeric_verify in TFLite converter.mlir_quantize() for quantization numeric verification. This pass adds NumericVerify ops for debugging during full-integer quantization, and NumericVerify op will output the differences between the original float & the dequantized values of the quantized values. The main characteristic of NumericVerify op is the output tensor name has the following format:
NumericVerify/{tensor_name}:{tensor_id}
where tensor_name and tensor_id are from the original quantized op's activation tensor (the first input tensor of the NumericVerify op) for the purpose of debugging. It provides users the easy way to search for tensors in the debugging tools.
If you turn on log_if_failed in NumericVerify op, then it will output logs & throw errors when there exist any errors > tolerance.
PiperOrigin-RevId: 348740984
Change-Id: Ifb965d2237e827a747d1888cbd21af311d42e4a1
ca3e7f55d9