How to convert a model (trained with quantization awareness) to int8?

I’ve being trying to convert a model trained with quantization awareness to a .tflite file that only uses int8 operations. I modified the quantization aware training example from guide adding the following lines before the conversion (as I would do in a usual conversion):

converter = tf.lite.TFLiteConverter.from_keras_model(q_aware_model)
converter.optimizations = [tf.lite.Optimize.OPTIMIZE_FOR_SIZE]

converter.representative_dataset = representative_data_gen

# Added the next 3 lines
converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.inference_input_type = tf.uint8
converter.inference_output_type = tf.uint8

quant_tflite_model = converter.convert()
file = open('quant_aware_model.tflite', 'wb')
file.write(quant_tflite_model)

However it raises the following error:

RuntimeError: Max and min for dynamic tensors should be recorded during calibration

I’ve checked the documentation but didn’t find any example of how to perform this operation. Is it supported?

[alanchiao]

If you think it’s too hard to find your answer in the above documentation and have thoughts, welcome any suggestions.

[EscVM]

@suttergustavo

I have the same problem! Have you found a solution?

I have a model trained with quantization aware. I followed the official TensorFlow guide, where is written “After this, you have an actually quantized model with int8 weights and uint8 activations.”

converter = tf.lite.TFLiteConverter.from_keras_model(q_aware_model) converter.optimizations = [tf.lite.Optimize.DEFAULT]

quantized_tflite_model = converter.convert()

The conversion works, but at the edge tpu complier says that the model is not quantized:

" Edge TPU Compiler version 2.1.302470888 Invalid model: bin/srmodel_full_integer_qaware.tflite Model not quantized "