def __init__(self,
min_init=-6,
max_init=6,
num_bits=8,
ema=False,
ema_decay=0.999,
per_channel=False,
out_channels=1,
quant_delay=0,
symmetric=False,
narrow_range=False):
"""init FakeQuantWithMinMax layer"""
super(FakeQuantWithMinMax, self).__init__()
self.min_init = min_init
self.num_bits = num_bits
self.max_init = max_init
self.ema = ema
self.ema_decay = ema_decay
self.per_channel = per_channel
self.out_channels = out_channels
self.quant_delay = quant_delay
self.symmetric = symmetric
self.narrow_range = narrow_range
if per_channel:
min_array = np.array([
self.min_init for i in range(0, self.out_channels)
]).astype(np.float32)
max_array = np.array([
self.max_init for i in range(0, self.channel_size)
]).astype(np.float32)
self.minq = Parameter(Tensor(min_array),
name='quant_min',
requires_grad=False)
self.maxq = Parameter(Tensor(max_array),
name='quant_max',
requires_grad=False)
self.fake_quant_train = P.FakeQuantWithMinMaxPerChannel(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=True)
self.fake_quant_infer = P.FakeQuantWithMinMaxPerChannel(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=False)
else:
min_array = np.array([min_init]).reshape(1).astype(np.float32)
max_array = np.array([max_init]).reshape(1).astype(np.float32)
self.minq = Parameter(Tensor(min_array),
name='quant_min',
requires_grad=False)
self.maxq = Parameter(Tensor(max_array),
name='quant_max',
requires_grad=False)
if context.get_context('device_target') == "Ascend":
self.fake_quant_train = FakeQuantWithMinMaxD(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=True,
min_init=self.minq,
max_init=self.maxq)
self.fake_quant_infer = FakeQuantWithMinMaxD(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=False,
min_init=self.minq,
max_init=self.maxq)
elif context.get_context('device_target') == "GPU":
self.fake_quant_train = P.FakeQuantWithMinMax(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=True)
self.fake_quant_infer = P.FakeQuantWithMinMax(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=ema_decay,
quant_delay=quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=False)
else:
raise ValueError("Not support platform.")
def __init__(self,
min_init=-6,
max_init=6,
num_bits=8,
ema=False,
ema_decay=0.999,
per_channel=False,
out_channels=1,
quant_delay=0,
symmetric=False,
narrow_range=False):
super(FakeQuantWithMinMax, self).__init__()
self.min_init = min_init
self.num_bits = num_bits
self.max_init = max_init
self.ema = ema
self.ema_decay = ema_decay
self.per_channel = per_channel
self.out_channels = out_channels
self.quant_delay = quant_delay
self.symmetric = symmetric
self.narrow_range = narrow_range
if per_channel:
min_array = np.array([
self.min_init for i in range(0, self.out_channels)
]).astype(np.float32)
max_array = np.array([
self.max_init for i in range(0, self.out_channels)
]).astype(np.float32)
self.fake_quant_train = P.FakeQuantWithMinMaxPerChannel(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=True)
self.fake_quant_infer = P.FakeQuantWithMinMaxPerChannel(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=False)
else:
min_array = np.array([min_init]).reshape(1).astype(np.float32)
max_array = np.array([max_init]).reshape(1).astype(np.float32)
self.fake_quant_train = P.FakeQuantWithMinMax(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=True)
self.fake_quant_infer = P.FakeQuantWithMinMax(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=False)
self.minq = Parameter(Tensor(min_array),
name='quant_min',
requires_grad=False)
self.maxq = Parameter(Tensor(max_array),
name='quant_max',
requires_grad=False)
def __init__(self,
min_init=-6,
max_init=6,
num_bits=8,
ema=False,
ema_decay=0.999,
per_channel=False,
channel_size=1,
quant_delay=0,
symmetric=False,
narrow_range=False,
training=True):
"""init FakeQuantWithMinMax ascend layer"""
super(FakeQuantWithMinMaxD, self).__init__()
self.min_init = min_init
self.num_bits = num_bits
self.max_init = max_init
self.ema = ema
self.ema_decay = ema_decay
self.per_channel = per_channel
self.channel_size = channel_size
self.quant_delay = quant_delay
self.symmetric = symmetric
self.narrow_range = narrow_range
self.training = training
if not per_channel:
self.fake_quant = P.FakeQuantWithMinMax(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=training)
self.ema_update = P.FakeQuantWithMinMaxUpdate(
num_bits=self.num_bits,
ema=self.ema,
ema_decay=self.ema_decay,
quant_delay=self.quant_delay,
symmetric=self.symmetric,
narrow_range=self.narrow_range,
training=training)
else:
raise RuntimeError("not support per channel")
if isinstance(min_init, Parameter):
self.minq = min_init
self.maxq = max_init
else:
self.minq = Parameter(Tensor(
np.array([min_init]).astype(np.float32)),
name='quant_min',
requires_grad=False)
self.maxq = Parameter(Tensor(
np.array([max_init]).astype(np.float32)),
name='quant_max',
requires_grad=False)
self.reduce_min = P.ReduceMin()
self.reduce_max = P.ReduceMax()