def __init__(self): super(QNN_HARnn, self).__init__() self.hardtanh0 = qnn.QuantHardTanh(quant_type=QuantType.INT, bit_width=2, narrow_range=True, bit_width_impl_type=BitWidthImplType.CONST, min_val = -1.0, max_val = 1.0, restrict_scaling_type=RestrictValueType.LOG_FP, scaling_per_channel=False, scaling_impl_type=ScalingImplType.PARAMETER) self.dropout0 = torch.nn.Dropout(p = DROPOUT) self.linear1 = qnn.QuantLinear(8, 128, bias=False, weight_quant_type=QuantType.BINARY, weight_bit_width=1, weight_scaling_stats_op = StatsOp.AVE, weight_scaling_stats_sigma=0.001, weight_scaling_per_output_channel = True, weight_narrow_range = True, weight_bit_width_impl_type=BitWidthImplType.CONST) self.hardtanh1 = qnn.QuantHardTanh(quant_type=QuantType.INT, bit_width=2, narrow_range=True, bit_width_impl_type=BitWidthImplType.CONST, min_val = -1.0, max_val = 1.0, restrict_scaling_type=RestrictValueType.LOG_FP, scaling_per_channel=False, scaling_impl_type=ScalingImplType.PARAMETER) self.dropout1 = torch.nn.Dropout(p = DROPOUT) self.linear2 = qnn.QuantLinear(128, 3, bias=False, weight_quant_type=QuantType.BINARY, weight_bit_width=1, weight_scaling_stats_op = StatsOp.AVE, weight_scaling_stats_sigma=0.001, weight_scaling_per_output_channel = False, weight_narrow_range = True, weight_bit_width_impl_type=BitWidthImplType.CONST)
def make_hardtanh_activation(bit_width, return_quant_tensor=False):
return quant_nn.QuantHardTanh(bit_width=bit_width,
max_val=ACT_MAX_VAL,
min_val=ACT_MIN_VAL,
quant_type=QUANT_TYPE,
scaling_impl_type=ACT_SCALING_IMPL_TYPE,
scaling_min_val=SCALING_MIN_VAL,
return_quant_tensor=return_quant_tensor)
def __init__(self, quant_type, bit_width, scaling_impl_type=ScalingImplType.CONST,
restrict_scaling_type=RestrictValueType.LOG_FP):
super(ZeroAct, self).__init__()
self._min_val_act = 0
self._max_val_act = 255
self._min_val = 0
self._max_val = 255
self._act = qnn.QuantHardTanh(scaling_impl_type=scaling_impl_type, restrict_scaling_type=restrict_scaling_type,
min_val=self._min_val_act, max_val=self._max_val_act, quant_type=quant_type,
bit_width=bit_width)
self._scale = (self._max_val - self._min_val) / (self._max_val_act - self._min_val_act)
def __init__(self, max_val, quant_type, bit_width, scaling_impl_type=ScalingImplType.CONST,
restrict_scaling_type=RestrictValueType.LOG_FP):
super(MyQuantReLU, self).__init__()
self._min_val_act = -1
self._max_val_act = 1 - 2 / (2 ** bit_width)
self._min_val = 0
self._max_val = max_val
self._act = qnn.QuantHardTanh(scaling_impl_type=scaling_impl_type, restrict_scaling_type=restrict_scaling_type,
min_val=self._min_val_act, max_val=self._max_val_act, quant_type=quant_type,
bit_width=bit_width)
self._scale = (self._max_val - self._min_val) / (self._max_val_act - self._min_val_act)
def make_quantization_input(bit_width, absolute_act_val, scaling_per_channel):
return quant_nn.QuantHardTanh(
bit_width=bit_width,
scaling_per_channel=scaling_per_channel,
quant_type=QUANT_TYPE,
scaling_impl_type=ACT_SCALING_IMPL_TYPE,
scaling_min_val=SCALING_MIN_VAL,
restrict_scaling_type=ACT_RESTRICT_SCALING_TYPE,
max_val=absolute_act_val,
min_val=-absolute_act_val,
return_quant_tensor=False)
def make_norm_scale(bit_width, absolute_act_val, scaling_per_channel):
return quant_nn.QuantHardTanh(
bit_width=bit_width,
scaling_per_channel=scaling_per_channel,
quant_type=QUANT_TYPE,
scaling_impl_type=ACT_SCALING_IMPL_TYPE,
scaling_min_val=SCALING_MIN_VAL,
restrict_scaling_type=ACT_RESTRICT_SCALING_TYPE,
max_val=absolute_act_val,
min_val=-absolute_act_val,
scaling_stats_permute_dims=(1, 0, 2),
return_quant_tensor=True)
def make_quant_hard_tanh(bit_width,
quant_type=QUANT_TYPE,
scaling_impl_type=ACT_SCALING_IMPL_TYPE,
scaling_per_channel=ACT_SCALING_PER_CHANNEL,
restrict_scaling_type=ACT_SCALING_RESTRICT_SCALING_TYPE,
scaling_min_val=SCALING_MIN_VAL,
threshold=HARD_TANH_THRESHOLD,
return_quant_tensor=ACT_RETURN_QUANT_TENSOR,
per_channel_broadcastable_shape=ACT_PER_CHANNEL_BROADCASTABLE_SHAPE):
return qnn.QuantHardTanh(bit_width=bit_width,
quant_type=quant_type,
scaling_per_channel=scaling_per_channel,
scaling_impl_type=scaling_impl_type,
restrict_scaling_type=restrict_scaling_type,
scaling_min_val=scaling_min_val,
max_val=threshold,
min_val=-threshold,
per_channel_broadcastable_shape=per_channel_broadcastable_shape,
return_quant_tensor=return_quant_tensor)
def make_quant_hard_tanh(bit_width,
scaling_impl_type = ACT_SCALING_IMPL_TYPE,
scaling_per_channel = ACT_SCALING_PER_CHANNEL,
restrict_scaling_type = ACT_SCALING_RESTRICT_SCALING_TYPE,
scaling_min_val = SCALING_MIN_VAL,
threshold = HARD_TANH_THRESHOLD,
min_overall_bit_width = MIN_OVERALL_BW,
max_overall_bit_width = MAX_OVERALL_BW,
return_quant_tensor = ACT_RETURN_QUANT_TENSOR,
per_channel_broadcastable_shape = ACT_PER_CHANNEL_BROADCASTABLE_SHAPE):
'''Helper for Hard Tanh activation layers'''
quant_type = get_quant_type(bit_width)
return qnn.QuantHardTanh(bit_width=bit_width,
quant_type=quant_type,
scaling_per_channel = scaling_per_channel,
scaling_impl_type = scaling_impl_type,
restrict_scaling_type = restrict_scaling_type,
scaling_min_val = scaling_min_val,
max_val = threshold,
min_val = -threshold,
min_overall_bit_width = min_overall_bit_width,
max_overall_bit_width = max_overall_bit_width,
per_channel_broadcastable_shape = per_channel_broadcastable_shape,
return_quant_tensor = return_quant_tensor)