def separable_conv_with_bn(x,
f,
stride=False,
aspp=False,
atrous_rate=1,
act_fn=True,
last_block=False,
end_point=False,
eps=1e-03,
out=False,
test=False,
fix_params=False):
with nn.parameter_scope("depthwise"):
if (stride == True):
h = PF.depthwise_convolution(x, (3, 3),
stride=(2, 2),
pad=(1, 1),
with_bias=False,
fix_parameters=fix_params)
elif (aspp == True):
h = PF.depthwise_convolution(x, (3, 3),
pad=(atrous_rate, atrous_rate),
stride=(1, 1),
dilation=(atrous_rate, atrous_rate),
with_bias=False,
fix_parameters=fix_params)
else:
h = PF.depthwise_convolution(x, (3, 3),
pad=(1, 1),
with_bias=False,
fix_parameters=fix_params)
h = PF.batch_normalization(h,
batch_stat=not test,
eps=eps,
fix_parameters=fix_params)
if last_block == True:
h = F.relu(h)
with nn.parameter_scope("pointwise"):
h = PF.convolution(h,
f, (1, 1),
stride=(1, 1),
with_bias=False,
fix_parameters=fix_params)
h = PF.batch_normalization(h,
batch_stat=not test,
eps=eps,
fix_parameters=fix_params)
if end_point == True:
global endpoints
endpoints['Decoder End Point 1'] = h
if act_fn == True:
h = F.relu(h)
return h
def shuffle_unit(x, scope_name, dn=False):
"""
Figure. 2 (b) and (c) in https://arxiv.org/pdf/1707.01083.pdf
"""
C = x.shape[1]
h = x
with nn.parameter_scope(scope_name):
with nn.parameter_scope("gconv1"):
h = PF.convolution(h, C, kernel=(1, 1), pad=(0, 0),
group=groups,
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
h = F.relu(h, True)
with nn.parameter_scope("shuffle"): # no meaning but semantics
h = shuffle(h)
with nn.parameter_scope("dconv"):
stride = (2, 2) if dn else (1, 1)
h = PF.depthwise_convolution(h, kernel=(3, 3), pad=(1, 1),
stride=stride,
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
with nn.parameter_scope("gconv2"):
h = PF.convolution(h, C, kernel=(1, 1), pad=(0, 0),
group=groups,
with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
s = F.average_pooling(x, (2, 2)) if dn else x
h = F.concatenate(*[h, s], axis=1) if dn else h + s
h = F.relu(h)
return h
def network(x, y, test=False):
# Input:x -> 3,64,64
# AveragePooling -> 3,12,21
h = F.average_pooling(x, (5, 3), (5, 3))
# LeakyReLU_2
h = F.leaky_relu(h, 0.1, True)
# Convolution_2 -> 20,13,21
h = PF.convolution(h, 20, (2, 3), (1, 1), name='Convolution_2')
# BatchNormalization
h = PF.batch_normalization(h, (1, ),
0.9,
0.0001,
not test,
name='BatchNormalization')
# ReLU
h = F.relu(h, True)
# DepthwiseConvolution
h = PF.depthwise_convolution(h, (5, 5), (2, 2),
name='DepthwiseConvolution')
# MaxPooling_2 -> 20,6,7
h = F.max_pooling(h, (2, 3), (2, 3))
# LeakyReLU
h = F.leaky_relu(h, 0.1, True)
# Affine -> 2
h = PF.affine(h, (2, ), name='Affine')
# Softmax
h = F.softmax(h)
return h
def test_FLOPsEstimator():
x = nn.Variable((1, 3, 12, 12))
y = PF.depthwise_convolution(x, kernel=(5, 5), with_bias=True)
t = PF.fused_batch_normalization(y)
z = F.relu6(F.sigmoid(PF.affine(t, (3, 3), base_axis=2) + 3))
z = F.global_average_pooling(z)
est = FLOPsEstimator()
assert est.predict(z) == 17644
def depthwise_separable_conv5x5(x, output_filter, scope, test):
"""
depthwise separable convolution with kernel 5x5.
"""
with nn.parameter_scope(scope):
h = conv1x1(x, output_filter, scope, test)
h = PF.depthwise_convolution(h, (5, 5), (2, 2), with_bias=False)
h = PF.convolution(h, output_filter, (1, 1), with_bias=False)
h = PF.batch_normalization(h, batch_stat=not test)
h = F.relu(h)
return h
def test_parametric_function_2d(inshape, kernel, multiplier, outshape):
base_axis = len(inshape) - 3
sample_channels = inshape[base_axis]
outmap_channels = sample_channels * multiplier
x = nn.Variable(inshape)
y = PF.depthwise_convolution(x, kernel, multiplier=multiplier)
p = nn.get_parameters()
assert y.shape == outshape
assert p['depthwise_conv/W'].shape == (outmap_channels,) + kernel
assert p['depthwise_conv/b'].shape == (outmap_channels,)
nn.clear_parameters()