How to create a representative dataset using TFrecords for tensorflow lite post training quantize conversion?

[Ethylbenzol]

1. System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Windows 10
• TensorFlow installation (pip package or built from source): pip
• TensorFlow library (version, if pip package or github SHA, if built from source): tf-nightly 2.6 dev

I am trying to create a representative data set for conversion to tflite and perform a int8 quantize operations. I was wondering, is it possible to use TFrecords as a input data to be used to create a representative_dataset instead of looping through each images to generate a representative_dataset? This is due to no examples of using TFrecords in the API but the API just says you can load TFrecord and create a dataset for manipulation.

[abattery]

@teijeong could you triage this issue?

[liufengdb]

Can you try the following code?

def representative_dataset_generator():
    """Dataset generator that generates random tensor with the same shape as the input"""
    filenames = [filename]
    raw_dataset = tf.data.TFRecordDataset(filenames) 
    for raw_record in raw_dataset.take(10):
       # manipulate the raw_record to get the right shape.
      yield raw_record    # raw_record is just a tensor

[Ethylbenzol]

Can you try the following code?

Just tried it. Threw a error TypeError: 'generator' object is not callable in tensorflow/tensorflow/lite/python/optimize/calibrator.py at:

    for sample in dataset_gen():
      if not initialized:
        initialized = True
        self._calibrator.Prepare([list(s.shape) for s in sample])
      self._calibrator.FeedTensor(sample)
    return self._calibrator.Calibrate()

I will try debug and see what went wrong with the conversion process.