I have been training a quantized Mobilenet V2 with TensorFlow, but I don't know how to get the class index from it.
I am using Tensorflow 1.12
Below are my input and output details.
Input details [{'name': 'normalized_input_image_tensor', 'index': 260, 'shape': array([ 1, 300, 300, 3], dtype=int32), 'shape_signature': array([ 1, 300, 300, 3], dtype=int32), 'dtype': <class 'numpy.uint8'>, 'quantization': (0.0078125, 128), 'quantization_parameters': {'scales': array([0.0078125], dtype=float32), 'zero_points': array([128], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
Output details [{'name': 'TFLite_Detection_PostProcess', 'index': 252, 'shape': array([ 1, 10, 4], dtype=int32), 'shape_signature': array([ 1, 10, 4], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'TFLite_Detection_PostProcess:1', 'index': 253, 'shape': array([ 1, 10], dtype=int32), 'shape_signature': array([ 1, 10], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'TFLite_Detection_PostProcess:2', 'index': 254, 'shape': array([ 1, 10], dtype=int32), 'shape_signature': array([ 1, 10], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}, {'name': 'TFLite_Detection_PostProcess:3', 'index': 255, 'shape': array([1], dtype=int32), 'shape_signature': array([1], dtype=int32), 'dtype': <class 'numpy.float32'>, 'quantization': (0.0, 0), 'quantization_parameters': {'scales': array([], dtype=float32), 'zero_points': array([], dtype=int32), 'quantized_dimension': 0}, 'sparsity_parameters': {}}]
I have been trying to get the class indices by doing the following:
interpreter = tf.lite.Interpreter(model_path=PATH)
interpreter.allocate_tensors()
input_details = interpreter.get_input_details()
output_details = interpreter.get_output_details()
interpreter.set_tensor(input_details[0]['index'], input_data)
interpreter.invoke()
classes = interpreter.get_tensor(output_details[1]['index'])[0]
However, the classes indices have been incorrect. When printed, classes looks like this: [0. 0. 0. 0. 0. 0. 0. 0. 0. 0.]. There are more than 1 classes in my dataset, so this doesn't make any sense.
What is the proper way to get the class indices?
After lots of experimentation, it turns out it was not a quantization issue. We were using the wrong graph_def .pb file when creating our .tflite, so it was predicting classes that did not exist.