[SOLVED] Using Saved Model Format from Tensorflow Lite Model Maker leads to poor prediction results

Using Saved Model Format from Tensorflow Lite Model Maker leads to poor prediction results

Following this tutorial I created an simple model for image classification with Tensorflow Model Maker. I changed export format from tflite to Saved Model as I intended to use it with tfjs-node as shown here.

# model.export(export_dir='.') 
model.export(export_dir='.', export_format=ExportFormat.SAVED_MODEL)

Testing the tflite model with an image in python returns very good results (near 95% for the desired class):

    import numpy as np
    from PIL import Image
    import tensorflow as tf
    
    model_file = "model.tflite"
    image_file = "my_image_name.jpg"
    
    interpreter = tf.lite.Interpreter(model_path=model_file)
    interpreter.allocate_tensors()
    
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
    
    image = Image.open(image_file)
    input_data = np.expand_dims(image, axis=0)
    
    interpreter.set_tensor(input_details[0]['index'], input_data)
    interpreter.invoke()
    output_data = interpreter.get_tensor(output_details[0]['index'])
    results = np.squeeze(output_data)
    for result in results:
        print(float(result / 255.0))
        # 0.011764705882352941 0.043137254901960784 0.9490196078431372

When using the Saved Model format prediction results are totally different and wrong:

    import numpy as np
    from PIL import Image
    from tensorflow.keras.models import load_model
    
    model_file = "saved_model"
    image_file = "my_image_name.jpg"
    
    model = load_model(model_file)    
    image = Image.open(image_file)
    input_tensor = np.expand_dims(image,axis=0)   
   
    predictions = model.predict(input_tensor)
    print(predictions)
    #0.0564433 0.69343865 0.25011805

Same problem when I use the Saved Model with tfjs-node.

1.) Do you have any idea how to fix this or where I could have a look at? 2.) What also could be helpful: Is there any supported way in tensorflow to use the tflite model with tfjs-node?

Things I tried among many other things:

Resizing Images to the desired size before creating the model to eliminate differences in image handling
Converting input_tensor to float32 doesn't make any difference (tf.convert_to_tensor(input_tensor, dtype=tf.float32))

Solution