def res_unit(x, scope_name, rng, dn=False, test=False):
C = x.shape[1]
with nn.parameter_scope(scope_name):
# Conv -> BN -> FixedPointQuantize -> Relu
with nn.parameter_scope("conv1"):
h = PF.fixed_point_quantized_convolution(x, C / 2, kernel=(1, 1), pad=(0, 0),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
h = F.fixed_point_quantize(h)
h = F.relu(h)
# Conv -> BN -> FixedPointQuantize -> Relu
with nn.parameter_scope("conv2"):
h = PF.fixed_point_quantized_convolution(h, C / 2, kernel=(3, 3), pad=(1, 1),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
h = F.fixed_point_quantize(h)
h = F.relu(h)
# Conv -> BN
with nn.parameter_scope("conv3"):
h = PF.fixed_point_quantized_convolution(h, C, kernel=(1, 1), pad=(0, 0),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
# Residual -> FixedPointQuantize -> Relu
h = F.fixed_point_quantize(h)
h = F.relu(h + x)
# Maxpooling
if dn:
h = F.max_pooling(h, kernel=(2, 2), stride=(2, 2))
return h
def fpq_weight_resblock(x,
maps,
kernel=(3, 3),
pad=(1, 1),
stride=(1, 1),
test=False,
sign=True,
n=8,
delta=2e-4,
name="convblock"):
h = x
with nn.parameter_scope(name):
h = PF.fixed_point_quantized_convolution(h,
maps,
kernel=kernel,
pad=pad,
stride=stride,
with_bias=False,
sign_w=sign,
n_w=n,
delta_w=delta,
sign_b=sign,
n_b=n,
delta_b=delta)
h = PF.batch_normalization(h, batch_stat=not test)
return F.relu(h + x)
def fpq_weight_lenet(image,
test=False,
n=8,
delta=2e-4,
name="fixed-point-weight-graph-ref"):
with nn.parameter_scope(name):
h = PF.fixed_point_quantized_convolution(image,
16, (5, 5), (1, 1),
with_bias=False,
n_w=n,
delta_w=delta,
name='conv1')
h = PF.batch_normalization(h, batch_stat=not test, name='conv1-bn')
h = F.max_pooling(h, (2, 2))
h = F.relu(h)
h = PF.fixed_point_quantized_convolution(h,
16, (5, 5), (1, 1),
with_bias=True,
n_w=n,
delta_w=delta,
name='conv2')
h = PF.batch_normalization(h, batch_stat=not test, name='conv2-bn')
h = F.max_pooling(h, (2, 2))
h = F.relu(h)
h = PF.fixed_point_quantized_affine(h,
10,
with_bias=False,
n_w=n,
delta_w=delta,
name='fc1')
h = PF.batch_normalization(h, batch_stat=not test, name='fc1-bn')
h = F.relu(h)
pred = PF.fixed_point_quantized_affine(h,
10,
with_bias=True,
n_w=n,
delta_w=delta,
name='fc2')
return pred
def _fixed_point_weight_conversion(self, inner_prod_func):
# Input
w_init = inner_prod_func.inputs[1].d
b_init = inner_prod_func.inputs[2].d if len(
inner_prod_func.inputs) == 3 else None
x = inner_prod_func.inputs[0]
x = self.input_map[x] if x in self.input_map else x
# Quantization params
sign_w = self.args_fpq["sign_w"]
n_w = self.args_fpq["n_w"]
delta_w = self.args_fpq["delta_w"]
sign_b = self.args_fpq["sign_b"]
n_b = self.args_fpq["n_b"]
quantize_b = self.args_fpq["quantize_b"]
delta_b = self.args_fpq["delta_b"]
# Determine delta
if not "delta_w" in self.args_fpq:
n = n_w - 1 if sign_w is True else n_w
w_abs_max = np.sort(np.abs(w_init.flatten()))[-1]
delta_w = 2**np.round_func(np.log2(w_abs_max / (2**n - 1)))
if not "delta_b" in self.args_fpq and len(inner_prod_func.inputs) == 3:
n = n_b - 1 if sign_b is True else n_b
b_abs_max = np.sort(np.abs(b_init.flatten()))[-1]
if b_abs_max != 0:
delta_b = 2**np.round_func(np.log2(b_abs_max / (2**n - 1)))
else:
delta_b = 0
# Parameter name
w = inner_prod_func.inputs[1]
idx = list(self.params.values()).index(w)
name = list(self.params.keys())[idx].rstrip("W/")
# Bias
with_bias = True if len(inner_prod_func.inputs) == 3 else False
# Conversion
if inner_prod_func.name == "Affine":
omaps = w_init.shape[1]
args = inner_prod_func.info.args
o = PF.fixed_point_quantized_affine(x,
omaps,
with_bias=with_bias,
w_init=w_init,
b_init=b_init,
sign_w=sign_w,
n_w=n_w,
delta_w=delta_w,
sign_b=sign_b,
n_b=n_b,
delta_b=delta_b,
name=name,
**args)
if inner_prod_func.name == "Convolution":
omaps = w_init.shape[0]
kernel = w_init.shape[2:]
args = inner_prod_func.info.args
if args.get('channel_last', False):
raise ValueError(
'channel_last=True is not supported in fixed_point_quantized_convolution.'
)
del args['channel_last']
o = PF.fixed_point_quantized_convolution(x,
omaps,
kernel,
with_bias=with_bias,
w_init=w_init,
b_init=b_init,
sign_w=sign_w,
n_w=n_w,
delta_w=delta_w,
sign_b=sign_b,
n_b=n_b,
delta_b=delta_b,
name=name,
**args)
if inner_prod_func.name == "Deconvolution":
omaps = w_init.shape[0]
kernel = w_init.shape[2:]
args = inner_prod_func.info.args
o = PF.fixed_point_quantized_deconvolution(x,
omaps,
kernel,
with_bias=with_bias,
w_init=w_init,
b_init=b_init,
sign_w=sign_w,
n_w=n_w,
delta_w=delta_w,
sign_b=sign_b,
n_b=n_b,
delta_b=delta_b,
name=name,
**args)
# Map output of ref graph to output of new graph
x = inner_prod_func.outputs[0]
self.input_map[x] = o
# Store output (just in case)
self.outputs.append(o)
return o
def cifar10_fp_net_resnet23_prediction(image, maps=64,
test=False):
"""
Construct Fixed-Point Net using resnet23. Fixed-Point Net quantizes weights
and activations using FixedPointQuantize function.
"""
# Residual Unit
def res_unit(x, scope_name, rng, dn=False, test=False):
C = x.shape[1]
with nn.parameter_scope(scope_name):
# Conv -> BN -> FixedPointQuantize -> Relu
with nn.parameter_scope("conv1"):
h = PF.fixed_point_quantized_convolution(x, C / 2, kernel=(1, 1), pad=(0, 0),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
h = F.fixed_point_quantize(h)
h = F.relu(h)
# Conv -> BN -> FixedPointQuantize -> Relu
with nn.parameter_scope("conv2"):
h = PF.fixed_point_quantized_convolution(h, C / 2, kernel=(3, 3), pad=(1, 1),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
h = F.fixed_point_quantize(h)
h = F.relu(h)
# Conv -> BN
with nn.parameter_scope("conv3"):
h = PF.fixed_point_quantized_convolution(h, C, kernel=(1, 1), pad=(0, 0),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
# Residual -> FixedPointQuantize -> Relu
h = F.fixed_point_quantize(h)
h = F.relu(h + x)
# Maxpooling
if dn:
h = F.max_pooling(h, kernel=(2, 2), stride=(2, 2))
return h
ncls = 10
# Conv -> BN -> FixedPointQuantize
with nn.parameter_scope("conv1"):
# Preprocess
image /= 255.0
if not test:
image = F.image_augmentation(image, contrast=1.0,
angle=0.25,
flip_lr=True)
image.need_grad = False
h = PF.fixed_point_quantized_convolution(image, maps, kernel=(3, 3), pad=(1, 1),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
h = F.fixed_point_quantize(h)
h = res_unit(h, "conv2", False) # -> 32x32
h = res_unit(h, "conv3", True) # -> 16x16
h = res_unit(h, "conv4", False) # -> 16x16
h = res_unit(h, "conv5", True) # -> 8x8
h = res_unit(h, "conv6", False) # -> 8x8
h = res_unit(h, "conv7", True) # -> 4x4
h = res_unit(h, "conv8", False) # -> 4x4
h = F.average_pooling(h, kernel=(4, 4)) # -> 1x1
pred = PF.fixed_point_quantized_affine(h, ncls)
return pred
def fpq_weight_small_resnet(image,
test=False,
sign=True,
n=8,
delta=2e-4,
name="fixed-point-graph-ref"):
with nn.parameter_scope(name):
h = image
h /= 255.0
h = PF.fixed_point_quantized_convolution(h,
16,
kernel=(3, 3),
pad=(1, 1),
sign_w=sign,
n_w=n,
delta_w=delta,
sign_b=sign,
n_b=n,
delta_b=delta,
with_bias=False,
name="first-conv")
h = PF.batch_normalization(h, name="first-bn", batch_stat=not test)
h = F.relu(h)
h = fpq_weight_resblock(h,
maps=16,
test=test,
sign=sign,
n=n,
delta=delta,
name="cb1")
h = fpq_weight_resblock(h,
maps=16,
test=test,
sign=sign,
n=n,
delta=delta,
name="cb2")
h = fpq_weight_resblock(h,
maps=16,
test=test,
sign=sign,
n=n,
delta=delta,
name="cb3")
h = fpq_weight_resblock(h,
maps=16,
test=test,
sign=sign,
n=n,
delta=delta,
name="cb4")
pred = PF.fixed_point_quantized_affine(h,
10,
sign_w=sign,
n_w=n,
delta_w=delta,
sign_b=sign,
n_b=n,
delta_b=delta,
name='fc')
return pred
