tflite – Model Inference

[中文]

The tflite module loads TensorFlow Lite .tflite flatbuffers and runs them through TensorFlow Lite Micro. It provides a general model execution interface for TFLite models, while higher-level task helpers can be layered on top when a model family needs shared preprocessing or post-processing.

Basic Usage

import tflite

model = tflite.Model("/sdcard/sine.tflite")
try:
    print("input:", model.input_shape, model.input_dtype, model.input_scale, model.input_zero_point)
    print("output:", model.output_shape, model.output_dtype, model.output_scale, model.output_zero_point)
    outputs = model.predict([input_array])
    raw = outputs[0]
finally:
    model.deinit()

predict() returns raw output tensors as ulab ndarrays. Quantized outputs are not automatically decoded; use output_scale and output_zero_point to convert values when the model requires it.

Callable Inputs

For image models it is often cheaper to fill the input tensor directly instead of building an intermediate ndarray:

def fill_input(buffer, shape, dtype_code):
    # Fill buffer with model-specific quantized bytes.
    ...

outputs = model.predict([fill_input])

The callable receives a writable bytearray view of the raw input tensor, the tensor shape, and an integer dtype code such as ord("b") for int8. This path is useful for image resizing, grayscale conversion, and quantization in Python examples.

Post-processing

Model can take a default postprocess callback, and predict() can take a one-shot callback. The callback receives (model, inputs, raw_outputs) and can return any Python object.

See also

Runnable examples: example/03-Machine-Learning/01-TFLite (person/no-person and sine smoke tests).

Classes

class tflite.Model(path, *, postprocess=...)

TensorFlow Lite Micro model wrapper.

Load a .tflite flatbuffer from the filesystem.

Parameters:

  • path – model path, for example “/sdcard/model.tflite” or “/flash/model.tflite”.

  • postprocess – optional callback that receives (model, inputs, raw_outputs) and returns the final result.

deinit()

Release model, interpreter, and tensor arena resources.

predict(inputs, *, callback=...)

Run inference.

Parameters:

  • inputs – one item per input tensor. Each item is an ulab ndarray or a callable that fills the raw tensor buffer.

  • callback – optional one-shot post-processing callback; overrides the default postprocess callback.