def res_unit(x, scope_name, rng, dn=False, test=False):
C = x.shape[1]
with nn.parameter_scope(scope_name):
# Conv -> BN -> Pow2Quantize -> Relu
with nn.parameter_scope("conv1"):
h = PF.pow2_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.pow2_quantize(h)
h = F.relu(h)
# Conv -> BN -> Pow2Quantize -> Relu
with nn.parameter_scope("conv2"):
h = PF.pow2_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.pow2_quantize(h)
h = F.relu(h)
# Conv -> BN
with nn.parameter_scope("conv3"):
h = PF.pow2_quantized_convolution(h, C, kernel=(1, 1), pad=(0, 0),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
# Residual -> Pow2Quantize -> Relu
h = F.pow2_quantize(h)
h = F.relu(h + x)
# Maxpooling
if dn:
h = F.max_pooling(h, kernel=(2, 2), stride=(2, 2))
return h
def cifar10_pow2_net_resnet23_prediction(image, maps=64,
test=False):
"""
Construct Pow2 Net using resnet23. Pow2 Net quantizes weights and
activations using Pow2Quantize 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 -> Pow2Quantize -> Relu
with nn.parameter_scope("conv1"):
h = PF.pow2_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.pow2_quantize(h)
h = F.relu(h)
# Conv -> BN -> Pow2Quantize -> Relu
with nn.parameter_scope("conv2"):
h = PF.pow2_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.pow2_quantize(h)
h = F.relu(h)
# Conv -> BN
with nn.parameter_scope("conv3"):
h = PF.pow2_quantized_convolution(h, C, kernel=(1, 1), pad=(0, 0),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
# Residual -> Pow2Quantize -> Relu
h = F.pow2_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 -> Pow2Quantize
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.pow2_quantized_convolution(image, maps, kernel=(3, 3), pad=(1, 1),
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
h = F.pow2_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.pow2_quantized_affine(h, ncls)
return pred