def __init__(self,
float_to_int_impl: Module,
bit_width_impl: Module,
quant_delay_steps: int = 0):
super(TruncIntQuant, self).__init__()
self.msb_clamp_bit_width_impl = bit_width_impl
self.float_to_int_impl = float_to_int_impl
self.delay_wrapper = DelayWrapper(quant_delay_steps)
def __init__(self,
scaling_impl: Module,
threshold: float,
quant_delay_steps: int = None):
super(TernaryQuant, self).__init__()
self.scaling_impl = scaling_impl
self.threshold = threshold
self.bit_width = BitWidthConst(2)
self.zero_point = StatelessBuffer(torch.tensor(0.0))
self.delay_wrapper = DelayWrapper(quant_delay_steps)
def __init__(
self,
scaling_impl: Module,
tensor_clamp_impl: Module = TensorClamp(),
quant_delay_steps: int = 0):
super(ClampedBinaryQuant, self).__init__()
self.scaling_impl = scaling_impl
self.bit_width = BitWidthConst(1)
self.zero_point = StatelessBuffer(torch.tensor(0.0))
self.delay_wrapper = DelayWrapper(quant_delay_steps)
self.tensor_clamp_impl = tensor_clamp_impl
def __init__(self,
narrow_range: bool,
signed: bool,
float_to_int_impl: Module = RoundSte(),
tensor_clamp_impl: Module = TensorClamp(),
quant_delay_steps: int = 0):
super(IntQuant, self).__init__()
self.float_to_int_impl = float_to_int_impl
self.tensor_clamp_impl = tensor_clamp_impl
self.signed = signed
self.narrow_range = narrow_range
self.delay_wrapper = DelayWrapper(quant_delay_steps)