Class ov::pass::ConvertQuantizeDequantize#

class ConvertQuantizeDequantize : public ov::pass::MatcherPass#

ConvertQuantizeDequantize transformation replaces following graph: FakeQuantize->Convert->Convert->Subtract->Multiply with a single FakeQuantize. Restrictions:

  • quantized data type must be i8 or u8

  • ’levels’ attribute to FakeQuantize must be equal to 256

  • (output_low, output_high) must be (-128, 127) or (0, 256) (depends on sign of quantized data type)

  • ’zero_point’ and ‘scale’ must be broadcastable to FakeQuantize’s output