How to expand dims of tensor for on device prediction?

General Discussion
, , ,
#1

Hello,

I’ve trained a custom image classification model using Keras, and I’m now trying to use it for prediction on an Android device (Kotlin). The model obviously takes as input a batch of tensors, so when predicting on Python on a single image, we need to use tf.expand_dims(image_tensor,0) to be able to input it, like so:

interpreter = tf.lite.Interpreter(TFLITE_FILE_PATH)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()

image = tf.io.decode_jpeg(tf.io.read_file(image_path))
image_without_alpha = image[:, :, 0:-1]
gray_scale_image = tf.image.rgb_to_grayscale(image_without_alpha)
resized_image = tf.image.resize(gray_scale_image, size=(224, 224))
input_data = tf.expand_dims(resized_image, 0)
interpreter.set_tensor(input_details['index'], input_data)
interpreter.invoke()
output_data = interpreter.get_tensor(output_details['index'])
print(np.argmax(output_data))

Yet, in Android, using TensorImage, it’s not clear how I can achieve this. I know it’s possible with TensorFlow Java, but it doesn’t seem to be available on Android.

Or am I supposed to add an extra step when converting the model to TFlite?

#3

Apparently ImageProcessor takes care of that, so you don’t need to do it manually. It’d be helpful if they explained that in the docs.