Why tflite api is different from regular tensorflow api

To perform inference on tflite model one has to write this code,

import numpy as np
import tensorflow as tf
interpreter = tf.lite.Interpreter(model_path=“converted_model.tflite”)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
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()
output_data = interpreter.get_tensor(output_details[0][‘index’])
print(output_data)

To run inference on regular tensorflow model, you will just do,

tfmodel.predict(test_dataset)

My question is why tflite api is not made simple and consistent with regular tensorflow model api? If they were same it would be very easy to remember and work with.

Also I was not able to find a function to evaluate tflite performance on test dataset, i.e.

tflite_model.evaluate(test_ds)

Am I missing something or there exist this method may be with a different name?

2 Likes

@markdaoust Do you have any historical background to share on why we don’t have a unified high level API?

I don’t have any answers for why.

But I do know that you can simplify your life quite a but using Interpreter.get_signature_list and get_signature_runner

Yes, we talked about this for a more articulate example at: