Missed TensorFlow World? Check out the recap. Learn more

TensorFlow Lite inference

The term inference refers to the process of executing a TensorFlow Lite model on-device in order to make predictions based on input data. To perform an inference with a TensorFlow Lite model, you must run it through an interpreter. The TensorFlow Lite interpreter is designed to be lean and fast. The interpreter uses a static graph ordering and a custom (less-dynamic) memory allocator to ensure minimal load, initialization, and execution latency.

This page describes how to access to the TensorFlow Lite interpreter and perform an inference using C++, Java, and Python, plus links to other resources for each supported platform.

Important concepts

TensorFlow Lite inference typically follows the following steps:

  1. Loading a model

    You must load the .tflite model into memory, which contains the model's execution graph.

  2. Transforming data

    Raw input data for the model generally does not match the input data format expected by the model. For example, you might need to resize an image or change the image format to be compatible with the model.

  3. Running inference

    This step involves using the TensorFlow Lite API to execute the model. It involves a few steps such as building the interpreter, and allocating tensors, as described in the following sections.

  4. Interpreting output

    When you receive results from the model inference, you must interpret the tensors in a meaningful way that's useful in your application.

    For example, a model might return only a list of probabilities. It's up to you to map the probabilities to relevant categories and present it to your end-user.

Supported platforms

TensorFlow inference APIs are provided for most common mobile/embedded platforms such as Android, iOS and Linux, in multiple programming languages.

In most cases, the API design reflects a preference for performance over ease of use. TensorFlow Lite is designed for fast inference on small devices, so it should be no surprise that the APIs try to avoid unnecessary copies at the expense of convenience. Similarly, consistency with TensorFlow APIs was not an explicit goal and some variance between languages is to be expected.

Across all libraries, the TensorFlow Lite API enables you to load models, feed inputs, and retrieve inference outputs.

Android

On Android, TensorFlow Lite inference can be performed using either Java or C++ APIs. The Java APIs provide convenience and can be used directly within your Android Activity classes. The C++ APIs offer more flexibility and speed, but may require writing JNI wrappers to move data between Java and C++ layers.

See below for details about using C++ and Java, or follow the Android quickstart for a tutorial and example code.

iOS

On iOS, TensorFlow Lite is available with native iOS libraries written in Swift and Objective-C.

This page doesn't include a discussion for about these languages, so you should refer to the iOS quickstart for a tutorial and example code.

Linux

On Linux platforms (including Raspberry Pi), you can run inferences using TensorFlow Lite APIs available in C++ and Python, as shown in the following sections.

Load and run a model in C++

Running a TensorFlow Lite model with C++ involves a few simple steps:

  1. Load the model into memory as a FlatBufferModel.
  2. Build an Interpreter based on an existing FlatBufferModel.
  3. Set input tensor values. (Optionally resize input tensors if the predefined sizes are not desired.)
  4. Invoke inference.
  5. Read output tensor values.

The FlatBufferModel class encapsulates a TensorFlow Lite model and you can build it in a couple of different ways, depending on where the model is stored:

class FlatBufferModel {
  // Build a model based on a file. Return a nullptr in case of failure.
  static std::unique_ptr<FlatBufferModel> BuildFromFile(
      const char* filename,
      ErrorReporter* error_reporter);

  // Build a model based on a pre-loaded flatbuffer. The caller retains
  // ownership of the buffer and should keep it alive until the returned object
  // is destroyed. Return a nullptr in case of failure.
  static std::unique_ptr<FlatBufferModel> BuildFromBuffer(
      const char* buffer,
      size_t buffer_size,
      ErrorReporter* error_reporter);
};

Now that you have the model as a FlatBufferModel object, you can execute it with an Interpreter. A single FlatBufferModel can be used simultaneously by more than one Interpreter.

The important parts of the Interpreter API are shown in the code snippet below. It should be noted that:

  • Tensors are represented by integers, in order to avoid string comparisons (and any fixed dependency on string libraries).
  • An interpreter must not be accessed from concurrent threads.
  • Memory allocation for input and output tensors must be triggered by calling AllocateTensors() right after resizing tensors.

The simplest usage of TensorFlow Lite with C++ looks like this:

// Load the model
std::unique_ptr<tflite::FlatBufferModel> model =
    tflite::FlatBufferModel::BuildFromFile(filename);

// Build the interpreter
tflite::ops::builtin::BuiltinOpResolver resolver;
std::unique_ptr<tflite::Interpreter> interpreter;
tflite::InterpreterBuilder(*model, resolver)(&interpreter);

// Resize input tensors, if desired.
interpreter->AllocateTensors();

float* input = interpreter->typed_input_tensor<float>(0);
// Fill `input`.

interpreter->Invoke();

float* output = interpreter->typed_output_tensor<float>(0);

For more example code, see minimal.cc and label_image.cc.

Load and run a model in Java

The Java API for running an inference with TensorFlow Lite is primarily designed for use with Android, so it's available as an Android library dependency: org.tensorflow:tensorflow-lite.

In Java, you'll use the Interpreter class to load a model and drive model inference. In many cases, this may be the only API you need.

You can initialize an Interpreter using a .tflite file:

public Interpreter(@NotNull File modelFile);

Or with a MappedByteBuffer:

public Interpreter(@NotNull MappedByteBuffer mappedByteBuffer);

In both cases, you must provide a valid TensorFlow Lite model or the API throws IllegalArgumentException. If you use MappedByteBuffer to initialize an Interpreter, it must remain unchanged for the whole lifetime of the Interpreter.

To then run an inference with the model, simply call Interpreter.run(). For example:

try (Interpreter interpreter = new Interpreter(file_of_a_tensorflowlite_model)) {
  interpreter.run(input, output);
}

The run() method takes only one input and returns only one output. So if your model has multiple inputs or multiple outputs, instead use:

interpreter.runForMultipleInputsOutputs(inputs, map_of_indices_to_outputs);

In this case, each entry in inputs corresponds to an input tensor and map_of_indices_to_outputs maps indices of output tensors to the corresponding output data.

In both cases, the tensor indices should correspond to the values you gave to the TensorFlow Lite Converter when you created the model. Be aware that the order of tensors in input must match the order given to the TensorFlow Lite Converter.

The Interpreter class also provides convenient functions for you to get the index of any model input or output using an operation name:

public int getInputIndex(String opName);
public int getOutputIndex(String opName);

If opName is not a valid operation in the model, it throws an IllegalArgumentException.

Also beware that Interpreter owns resources. To avoid memory leak, the resources must be released after use by:

interpreter.close();

For an example project with Java, see the Android image classification sample.

Supported data types (in Java)

To use TensorFlow Lite, the data types of the input and output tensors must be one of the following primitive types:

  • float
  • int
  • long
  • byte

String types are also supported, but they are encoded differently than the primitive types. In particular, the shape of a string Tensor dictates the number and arrangement of strings in the Tensor, with each element itself being a variable length string. In this sense, the (byte) size of the Tensor cannot be computed from the shape and type alone, and consequently strings cannot be provided as a single, flat ByteBuffer argument.

If other data types, including boxed types like Integer and Float, are used, an IllegalArgumentException will be thrown.

Inputs

Each input should be an array or multi-dimensional array of the supported primitive types, or a raw ByteBuffer of the appropriate size. If the input is an array or multi-dimensional array, the associated input tensor will be implicitly resized to the array's dimensions at inference time. If the input is a ByteBuffer, the caller should first manually resize the associated input tensor (via Interpreter.resizeInput()) before running inference.

When using ByteBuffer, prefer using direct byte buffers, as this allows the Interpreter to avoid unnecessary copies. If the ByteBuffer is a direct byte buffer, its order must be ByteOrder.nativeOrder(). After it is used for a model inference, it must remain unchanged until the model inference is finished.

Outputs

Each output should be an array or multi-dimensional array of the supported primitive types, or a ByteBuffer of the appropriate size. Note that some models have dynamic outputs, where the shape of output tensors can vary depending on the input. There's no straightforward way of handling this with the existing Java inference API, but planned extensions will make this possible.

Load and run a model in Python

The Python API for running an inference is provided in the tf.lite module. From which, you mostly need only tf.lite.Interpreter to load a model and run an inference.

The following example shows how to use the Python interpreter to load a .tflite file and run inference with random input data:

import numpy as np
import tensorflow as tf

# Load TFLite model and allocate tensors.
interpreter = tf.lite.Interpreter(model_path="converted_model.tflite")
interpreter.allocate_tensors()

# Get input and output tensors.
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

# Test model on random input data.
input_shape = input_details[0]['shape']
input_data = np.array(np.random.random_sample(input_shape), dtype=np.float32)
interpreter.set_tensor(input_details[0]['index'], input_data)

interpreter.invoke()

# The function `get_tensor()` returns a copy of the tensor data.
# Use `tensor()` in order to get a pointer to the tensor.
output_data = interpreter.get_tensor(output_details[0]['index'])
print(output_data)

Alternative to loading the model as a pre-converted .tflite file, you can combine your code with the TensorFlow Lite Converter Python API (tf.lite.TFLiteConverter), allowing you to convert your TensorFlow model into the TensorFlow Lite format and then run an inference:

import numpy as np
import tensorflow as tf

img = tf.placeholder(name="img", dtype=tf.float32, shape=(1, 64, 64, 3))
const = tf.constant([1., 2., 3.]) + tf.constant([1., 4., 4.])
val = img + const
out = tf.identity(val, name="out")

# Convert to TF Lite format
with tf.Session() as sess:
  converter = tf.lite.TFLiteConverter.from_session(sess, [img], [out])
  tflite_model = converter.convert()

# Load TFLite model and allocate tensors.
interpreter = tf.lite.Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()

# Continue to get tensors and so forth, as shown above...

For more Python sample code, see label_image.py.

Tip: Run help(tf.lite.Interpreter) in the Python terminal to get detailed documentation about the interpreter.

Write a custom operator

All TensorFlow Lite operators (both custom and builtin) are defined using a simple pure-C interface that consists of four functions:

typedef struct {
  void* (*init)(TfLiteContext* context, const char* buffer, size_t length);
  void (*free)(TfLiteContext* context, void* buffer);
  TfLiteStatus (*prepare)(TfLiteContext* context, TfLiteNode* node);
  TfLiteStatus (*invoke)(TfLiteContext* context, TfLiteNode* node);
} TfLiteRegistration;

Refer to context.h for details on TfLiteContext and TfLiteNode. The former provides error reporting facilities and access to global objects, including all the tensors. The latter allows implementations to access their inputs and outputs.

When the interpreter loads a model, it calls init() once for each node in the graph. A given init() will be called more than once if the op is used multiple times in the graph. For custom ops a configuration buffer will be provided, containing a flexbuffer that maps parameter names to their values. The buffer is empty for builtin ops because the interpreter has already parsed the op parameters. Kernel implementation that require state should initialize it here and transfer ownership to the caller. For each init() call, there will be a corresponding call to free(), allowing implementations to dispose of the buffer they might have allocated in init().

Whenever the input tensors are resized the interpreter will go through the graph notifying implementations of the change. This gives them the chance to resize their internal buffer, check validity of input shapes and types, and recalculate output shapes. This is all done through prepare() and implementation can access their state using node->user_data.

Finally, each time inference runs the interpreter traverses the graph calling invoke(), and here too the state is available as node->user_data.

Custom ops can be implemented in exactly the same way as builtin ops, by defined those four functions and a global registration function that usually looks like this:

namespace tflite {
namespace ops {
namespace custom {
  TfLiteRegistration* Register_MY_CUSTOM_OP() {
    static TfLiteRegistration r = {my_custom_op::Init,
                                   my_custom_op::Free,
                                   my_custom_op::Prepare,
                                   my_custom_op::Eval};
    return &r;
  }
}  // namespace custom
}  // namespace ops
}  // namespace tflite

Note that registration is not automatic and an explicit call to Register_MY_CUSTOM_OP should be made somewhere. While the standard BuiltinOpResolver (available from the :builtin_ops target) takes care of the registration of builtins, custom ops will have to be collected in separate custom libraries.

Customize the kernel library

Behind the scenes the interpreter will load a library of kernels which will be assigned to execute each of the operators in the model. While the default library only contains builtin kernels, it is possible to replace it with a custom library.

The interpreter uses an OpResolver to translate operator codes and names into actual code:

class OpResolver {
  virtual TfLiteRegistration* FindOp(tflite::BuiltinOperator op) const = 0;
  virtual TfLiteRegistration* FindOp(const char* op) const = 0;
  virtual void AddOp(tflite::BuiltinOperator op, TfLiteRegistration* registration) = 0;
  virtual void AddOp(const char* op, TfLiteRegistration* registration) = 0;
};

Regular usage requires that you use the BuiltinOpResolver and write:

tflite::ops::builtin::BuiltinOpResolver resolver;

You can optionally register custom ops (before you pass the resolver to the InterpreterBuilder):

resolver.AddOp("MY_CUSTOM_OP", Register_MY_CUSTOM_OP());

If the set of builtin ops is deemed to be too large, a new OpResolver could be code-generated based on a given subset of ops, possibly only the ones contained in a given model. This is the equivalent of TensorFlow's selective registration (and a simple version of it is available in the tools directory).