How test the accuracy for tensorflow lite model

Question

I train a CNN tensorflow model, and convert for a tensorflow lite model. And now i want to know how can i make the evaluate for the TFLITE model.

I make this code :

interpreter = tf.lite.Interpreter(model_save)
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details() 


interpreter.resize_tensor_input(input_details[0]['index'], ((len(X_test)), 180,180, 3))
interpreter.resize_tensor_input(output_details[0]['index'], (len(X_test), 4))
interpreter.allocate_tensors()

input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

interpreter.set_tensor(input_details[0]['index'], X_test)
interpreter.invoke()

loss, accuracy = interpreter.evaluate(X_test)

But, show me the error:

'Interpreter' object has no attribute 'evaluate'

After this I tried:

loss, accuracy = interpreter.evaluate_tflite(X_test)

But apparently this just it works for Model Makers model. So now i just don't know how to preceed.

Answer 1

If your tflite model has a signature, then you can leverage the signature, see the guide.

If the model doesn't have signatures, then you can see what are the outputs like this

output_details = interpreter.get_output_details()
# Assuming you have 2 outputs
output_1 = interpreter.get_tensor(output_details[0]['index'])
output_2 = interpreter.get_tensor(output_details[1]['index'])

The signature is the suggested way, and is safer for output reordering issues.