Closed
Description
1. System information
- OS Platform and Distribution (e.g., Linux Ubuntu 16.04): 18.04
- TensorFlow installation (pip package or built from source): pip
- TensorFlow library (version, if pip package or github SHA, if built from source): 2.5.0
2. Code
Provide code to help us reproduce your issues using one of the following options:
import tensorflow as tf
import numpy as np
keras = tf.keras
layers = keras.layers
def infer(tflite_model, img):
interpreter = tf.lite.Interpreter(model_content=tflite_model)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()[0]
interpreter.set_tensor(input_details["index"], img)
interpreter.invoke()
output_details = interpreter.get_output_details()[0]
output = interpreter.get_tensor(output_details["index"])
return output
def convert_to_tflite(model):
converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.experimental_new_converter = True
tflite_model = converter.convert()
return tflite_model
# Image to test
np.random.seed(1)
input_shape = (12, 14, 18)
img = np.random.random((1,)+input_shape).astype(np.float32) * 7.0 - 3.5
# Create a model
i = layers.Input(shape=input_shape)
x = tf.quantization.fake_quant_with_min_max_args(i, min=0.0, max=3.984375)
x = tf.reduce_sum(x, axis=None)
x = tf.quantization.fake_quant_with_min_max_args(x, min=-4.0, max=3.96875)
model = keras.models.Model(inputs=i, outputs=x)
tf_output = model.predict(img)
tflite_model = convert_to_tflite(model)
tflite_output = infer(tflite_model, img)
print(f'Tensorflow output: {tf_output}, TFLite output: {tflite_output}, ')3. Failure after conversion
If the conversion is successful, but the generated model is wrong, then state what is wrong:
I get a huge difference between the TF model and the TFLite model:
Tensorflow output: 3.96875, TFLite output: -2.78125,
The Tensorflow model gives completely different results than the TFLite model. Furthermore, before the reduce_sum, I quantize the tensor using only positive values, so it does not make sense that the output of the model is negative (which is what I get).