Thanks for tuning in to Google I/O. View all sessions on demandWatch on demand

tflite

\file

Summary

For documentation, see tensorflow/lite/core/interpreter.h.

Memory management for TF Lite.

This provides a few C++ helpers that are useful for manipulating C structures in C++.

Main abstraction controlling the tflite interpreter. Do NOT include this file directly, instead include third_party/tensorflow/lite/interpreter.h See third_party/tensorflow/lite/c/common.h for the API for defining operations (TfLiteRegistration).

Provides functionality to construct an interpreter for a model.

WARNING: Users of TensorFlow Lite should not include this file directly, but should instead include "third_party/tensorflow/lite/interpreter_builder.h". Only the TensorFlow Lite implementation itself should include this file directly.

Deserialization infrastructure for tflite. Provides functionality to go from a serialized tflite model in flatbuffer format to an in-memory representation of the model.

WARNING: Users of TensorFlow Lite should not include this file directly, but should instead include "third_party/tensorflow/lite/model_builder.h". Only the TensorFlow Lite implementation itself should include this file directly.

Typedefs
FlatBufferModel using
impl::FlatBufferModel
Interpreter typedef
::tflite::impl::Interpreter
An interpreter for a graph of nodes that input and output from tensors.
InterpreterBuilder using
impl::InterpreterBuilder
Build an interpreter capable of interpreting model.

Functions
DefaultErrorReporter()
GetRegistrationFromOpCode(const OperatorCode *opcode, const OpResolver & op_resolver, ErrorReporter *error_reporter, const TfLiteRegistration **registration)
TfLiteStatus

Classes
tflite::Allocation

A memory allocation handle. This could be a mmap or shared memory.

tflite::ErrorReporter

A functor that reports error to supporting system.

tflite::FileCopyAllocation
tflite::MMAPAllocation

Note that not all platforms support MMAP-based allocation.

tflite::MemoryAllocation
tflite::MutableOpResolver

An OpResolver that is mutable, also used as the op in gen_op_registration.

tflite::OpResolver

Abstract interface that returns TfLiteRegistrations given op codes or custom op names.

tflite::TfLiteIntArrayView

Provides a range iterable wrapper for TfLiteIntArray* (C lists) that TfLite C api uses.

Structs
tflite::StderrReporter

Namespaces
tflite::impl

An RAII object that represents a read-only tflite model, copied from disk, or mmapped.

tflite::op_resolver_hasher

Typedefs

FlatBufferModel

impl::FlatBufferModel FlatBufferModel

Interpreter

::tflite::impl::Interpreter Interpreter

An interpreter for a graph of nodes that input and output from tensors.

Each node of the graph processes a set of input tensors and produces a set of output Tensors. All inputs/output tensors are referenced by index.

Usage:


// Create model from file. Note that the model instance must outlive the
// interpreter instance.
auto model = tflite::FlatBufferModel::BuildFromFile(...);
if (model == nullptr) {
  // Return error.
}
// Create an Interpreter with an InterpreterBuilder.
std::unique_ptr interpreter;
tflite::ops::builtin::BuiltinOpResolver resolver;
if (InterpreterBuilder(*model, resolver)(&interpreter) != kTfLiteOk) {
  // Return failure.
}
if (interpreter->AllocateTensors() != kTfLiteOk) {
  // Return failure.
}

auto input = interpreter->typed_tensor(0);
for (int i = 0; i < input_size; i++) {
  input[i] = ...;  interpreter->Invoke();

Note: For nearly all practical use cases, one should not directly construct an Interpreter object, but rather use the InterpreterBuilder.

\warning This class is not thread-safe. The client is responsible for ensuring serialized interaction to avoid data races and undefined behavior.

InterpreterBuilder

impl::InterpreterBuilder InterpreterBuilder

Build an interpreter capable of interpreting model.

  • model: A model whose lifetime must be at least as long as any interpreter(s) created by the builder. In principle multiple interpreters can be made from a single model.
  • op_resolver: An instance that implements the OpResolver interface, which maps custom op names and builtin op codes to op registrations. The lifetime of the provided op_resolver object must be at least as long as the InterpreterBuilder; unlike model and error_reporter, the op_resolver does not need to exist for the duration of any created Interpreter objects.
  • error_reporter: a functor that is called to report errors that handles printf var arg semantics. The lifetime of the error_reporter object must be greater than or equal to the Interpreter created by operator().
  • options_experimental: Options that can change behavior of interpreter. WARNING: this parameter is an experimental API and is subject to change.

Returns a kTfLiteOk when successful and sets interpreter to a valid Interpreter. Note: The user must ensure the lifetime of the model (and error reporter, if provided) is at least as long as interpreter's lifetime, and a single model instance may safely be used with multiple interpreters.

Functions

DefaultErrorReporter

ErrorReporter * DefaultErrorReporter()

GetRegistrationFromOpCode

TfLiteStatus GetRegistrationFromOpCode(
  const OperatorCode *opcode,
  const OpResolver & op_resolver,
  ErrorReporter *error_reporter,
  const TfLiteRegistration **registration
)