def __init__(self,
kernel_size: int,
stride: int = None,
signed: bool = True,
min_overall_bit_width: Optional[int] = 2,
max_overall_bit_width: Optional[int] = 32,
quant_type: QuantType = QuantType.FP,
lsb_trunc_bit_width_impl_type=BitWidthImplType.CONST):
QuantLayer.__init__(self,
compute_output_scale=True,
compute_output_bit_width=True,
return_quant_tensor=True)
AvgPool2d.__init__(self, kernel_size=kernel_size, stride=stride)
ls_bit_width_to_trunc = math.ceil(math.log2(kernel_size * kernel_size))
self.signed = signed
self.quant_type = quant_type
explicit_rescaling = True # we are explicitly rescaling as we are replacing the div in avg with trunc
self.accumulator_quant = TruncQuantProxy(
signed=signed,
quant_type=quant_type,
trunc_at_least_init_val=True,
ls_bit_width_to_trunc=ls_bit_width_to_trunc,
min_overall_bit_width=min_overall_bit_width,
max_overall_bit_width=max_overall_bit_width,
lsb_trunc_bit_width_impl_type=lsb_trunc_bit_width_impl_type,
explicit_rescaling=explicit_rescaling,
override_pretrained_bit_width=False)
def __init__(
self,
in_channels: int,
out_channels: int,
kernel_size: Union[int, Tuple[int]],
stride: Union[int, Tuple[int]] = 1,
padding: Union[int, Tuple[int]] = 0,
output_padding: Union[int, Tuple[int]] = 0,
padding_type: PaddingType = PaddingType.STANDARD,
dilation: Union[int, Tuple[int]] = 1,
groups: int = 1,
bias: bool = True,
bias_quant_type: QuantType = QuantType.FP,
bias_narrow_range: bool = False,
bias_bit_width: int = None,
weight_quant_override: WeightQuantProxy = None,
weight_quant_type: QuantType = QuantType.FP,
weight_narrow_range: bool = False,
weight_scaling_override: Optional[Module] = None,
weight_bit_width_impl_override: Union[BitWidthParameter,
BitWidthConst] = None,
weight_bit_width_impl_type: BitWidthImplType = BitWidthImplType.
CONST,
weight_restrict_bit_width_type:
RestrictValueType = RestrictValueType.INT,
weight_bit_width: int = 32,
weight_min_overall_bit_width: Optional[int] = 2,
weight_max_overall_bit_width: Optional[int] = None,
weight_scaling_impl_type: ScalingImplType = ScalingImplType.STATS,
weight_scaling_const: Optional[float] = None,
weight_scaling_stats_op: StatsOp = StatsOp.MAX,
weight_scaling_per_output_channel: bool = False,
weight_ternary_threshold: float = 0.5,
weight_restrict_scaling_type: RestrictValueType = RestrictValueType
.LOG_FP,
weight_scaling_stats_sigma: float = 3.0,
weight_scaling_min_val: float = SCALING_MIN_VAL,
weight_override_pretrained_bit_width: bool = False,
compute_output_scale: bool = False,
compute_output_bit_width: bool = False,
return_quant_tensor: bool = False,
deterministic: bool = False) -> None:
QuantLayer.__init__(self,
compute_output_scale=compute_output_scale,
compute_output_bit_width=compute_output_bit_width,
return_quant_tensor=return_quant_tensor)
ConvTranspose1d.__init__(self,
in_channels=in_channels,
out_channels=out_channels,
kernel_size=kernel_size,
stride=stride,
padding=padding,
output_padding=output_padding,
dilation=dilation,
groups=groups,
bias=bias)
if weight_quant_type == QuantType.FP and compute_output_bit_width:
raise Exception(
"Computing output bit width requires enabling quantization")
if bias_quant_type != QuantType.FP and not (compute_output_scale and
compute_output_bit_width):
raise Exception(
"Quantizing bias requires to compute output scale and output bit width"
)
if torch.backends.cudnn.benchmark:
torch.backends.cudnn.deterministic = deterministic
# self.per_elem_ops = 2 * self.kernel_size[0] * (in_channels // groups) # TO DO: Implement op_count
self.padding_type = padding_type
self.weight_reg = WeightReg()
if weight_quant_override is not None:
self.weight_quant = weight_quant_override
self.weight_quant.add_tracked_parameter(self.weight)
else:
weight_scaling_stats_input_concat_dim = 1
if weight_scaling_per_output_channel:
weight_stats_input_view_shape_impl = StatsInputViewShapeImpl.OVER_OUTPUT_CHANNELS
weight_scaling_shape = self.per_output_channel_broadcastable_shape
weight_scaling_stats_reduce_dim = 1
else:
weight_stats_input_view_shape_impl = StatsInputViewShapeImpl.OVER_TENSOR
weight_scaling_shape = SCALING_SCALAR_SHAPE
weight_scaling_stats_reduce_dim = None
if weight_scaling_stats_op == StatsOp.MAX_AVE:
weight_stats_input_view_shape_impl = StatsInputViewShapeImpl.OVER_OUTPUT_CHANNELS
weight_scaling_stats_reduce_dim = 1
self.weight_quant = WeightQuantProxy(
bit_width=weight_bit_width,
quant_type=weight_quant_type,
narrow_range=weight_narrow_range,
scaling_override=weight_scaling_override,
restrict_scaling_type=weight_restrict_scaling_type,
scaling_const=weight_scaling_const,
scaling_stats_op=weight_scaling_stats_op,
scaling_impl_type=weight_scaling_impl_type,
scaling_stats_reduce_dim=weight_scaling_stats_reduce_dim,
scaling_shape=weight_scaling_shape,
bit_width_impl_type=weight_bit_width_impl_type,
bit_width_impl_override=weight_bit_width_impl_override,
restrict_bit_width_type=weight_restrict_bit_width_type,
min_overall_bit_width=weight_min_overall_bit_width,
max_overall_bit_width=weight_max_overall_bit_width,
tracked_parameter_list_init=self.weight,
ternary_threshold=weight_ternary_threshold,
scaling_stats_input_view_shape_impl=
weight_stats_input_view_shape_impl,
scaling_stats_input_concat_dim=
weight_scaling_stats_input_concat_dim,
scaling_stats_sigma=weight_scaling_stats_sigma,
scaling_min_val=weight_scaling_min_val,
override_pretrained_bit_width=
weight_override_pretrained_bit_width)
self.bias_quant = BiasQuantProxy(quant_type=bias_quant_type,
bit_width=bias_bit_width,
narrow_range=bias_narrow_range)