tflite/guide/inference.md

The model is created by reinterpret a raw memory directly.

call graph

Interpreter::Invoke
Subgraph::Invoke
Subgraph::OpInvoke
TfLiteRegistration::Invoke

Op Registration

In tflite, ops are registerred to the interpreter.

1. lite/python/optimize/calibrator_wrapper.cc:CreateWrapperCPPFromBuffer > lite/kernels/register.cc:BuiltinOpResolver
   1. BuiltinOpResolver creates a default registration
2. the Resolver is used by BuildLocalIndexToRegistrationMapping
   1. which create flatbuffer_op_index_to_registration_
3. The mapping is used by model.cc:InterpreterBuilder::ParseNodes
   1. which retrieve registration from op_index
   2. the registration and node is added by Subgraph::AddNodeWithParameter
4. The node and registration is used in Invoke

when is minmax being updated?

1. the info is stored in lite/schema :TensorT:quantization
2. moved by lite/tools/optimize/calibration/calibration_reader.cc:AddCalibrationToModel
   1. from logger_ -> GetCalibrationValues
3. created in lite/tools/optimize/calibration/calibrator.cc:BuildLoggingInterpreter
   1. in GetCalibratorRegistry()->CreateCalibrator
4. Calibrator is based on result stored in LoggingOpResolver
5. LoggingOpResolver overwrite TfLiteRegistration::invoke
   1. with calibrator.cc:LoggingEval
   2. calibrator is retrieved from the context. (singleton)
6. calls LogTensorValue
   1. updates tensor stats map calibration_logger.h:Update
   2. with min, max -> absolute value

how does tensor get indexed by int?

in tensorflow, tensor is refereced by string

todo: how does this work in tflite

special ops

add

scale and zero point

Reference

calculation

Reference

concat

scale and zero point is taken as an array

Reference

if input and output scale is the same, memcpy if they are different,

Reference