def inception_17(hidden_channel, name):
# 1x1
s1 = ConvBN(768, 192, 1, 1, 0, name['1x1'][0])
# 7x7
s21 = ConvBN(768, hidden_channel, 1, 1, 0, name['7x7'][0])
s22 = ConvBN(hidden_channel, hidden_channel, (1, 7), (1, 1),
(0, 3), name['7x7'][1])
s23 = ConvBN(hidden_channel, 192, (7, 1), (1, 1), (3, 0),
name['7x7'][2])
s2 = Sequential(s21, s22, s23)
# double 7x7
s31 = ConvBN(768, hidden_channel, 1, 1, 0, name['double7x7'][0])
s32 = ConvBN(hidden_channel, hidden_channel, (7, 1), (1, 1),
(3, 0), name['double7x7'][1])
s33 = ConvBN(hidden_channel, hidden_channel, (1, 7), (1, 1),
(0, 3), name['double7x7'][2])
s34 = ConvBN(hidden_channel, hidden_channel, (7, 1), (1, 1),
(3, 0), name['double7x7'][3])
s35 = ConvBN(hidden_channel, 192, (1, 7), (1, 1), (0, 3),
name['double7x7'][4])
s3 = Sequential(s31, s32, s33, s34, s35)
# pool
s4 = ConvBN(768,
192,
1,
1,
0,
name['pool'][1],
pool=A.AveragePooling2D(3, 1, 1))
return Inception(s1, s2, s3, s4)
def inception_8(input_channel, name):
# 1x1
s1 = ConvBN(input_channel, 320, 1, 1, 0, name['1x1'][0])
# 3x3
s21 = ConvBN(input_channel, 384, 1, 1, 0, name['3x3'][0])
s22 = Inception(
ConvBN(384, 384, (1, 3), (1, 1), (0, 1), name['3x3'][1]),
ConvBN(384, 384, (3, 1), (1, 1), (1, 0), name['3x3'][2]))
s2 = Sequential(s21, s22)
# double 3x3
s31 = ConvBN(input_channel, 448, 1, 1, 0, name['double3x3'][0])
s32 = ConvBN(448, 384, 3, 1, 1, name['double3x3'][1])
s331 = ConvBN(384, 384, (1, 3), (1, 1), (0, 1),
name['double3x3'][2])
s332 = ConvBN(384, 384, (3, 1), (1, 1), (1, 0),
name['double3x3'][3])
s33 = Inception(s331, s332)
s3 = Sequential(s31, s32, s33)
# pool
s4 = ConvBN(input_channel,
192,
1,
1,
0,
name['pool'][1],
pool=A.AveragePooling2D(3, 1, 1))
return Inception(s1, s2, s3, s4)
def inception_1(hidden_channel):
# 1x1
s1 = AuxConv(C.Convolution2D(768, 192, 1, use_cudnn=use_cudnn))
# 7x7
s21 = AuxConv(
C.Convolution2D(768, hidden_channel, 1, use_cudnn=use_cudnn))
s22 = AuxConv(
C.Convolution2D(hidden_channel,
hidden_channel, (1, 7),
pad=(0, 3),
use_cudnn=use_cudnn))
s23 = AuxConv(
C.Convolution2D(hidden_channel,
192, (7, 1),
pad=(3, 0),
use_cudnn=use_cudnn))
s2 = Sequential(s21, s22, s23)
# double 7x7
s31 = AuxConv(
C.Convolution2D(768, hidden_channel, 1, use_cudnn=use_cudnn))
s32 = AuxConv(
C.Convolution2D(hidden_channel,
hidden_channel, (1, 7),
pad=(0, 3),
use_cudnn=use_cudnn))
s33 = AuxConv(
C.Convolution2D(hidden_channel,
hidden_channel, (7, 1),
pad=(3, 0),
use_cudnn=use_cudnn))
s34 = AuxConv(
C.Convolution2D(hidden_channel,
hidden_channel, (1, 7),
pad=(0, 3),
use_cudnn=use_cudnn))
s35 = AuxConv(
C.Convolution2D(hidden_channel,
192, (7, 1),
pad=(3, 0),
use_cudnn=use_cudnn))
s3 = Sequential(s31, s32, s33, s34, s35)
# pool
s4 = AuxConv(C.Convolution2D(768,
192,
3,
pad=1,
use_cudnn=use_cudnn),
pool=A.AveragePooling2D(3, 1, 1, use_cudnn=use_cudnn))
return Inception(s1, s2, s3, s4)
def inception_2(input_channel):
# 1x1
s1 = AuxConv(
C.Convolution2D(input_channel, 320, 1, use_cudnn=use_cudnn))
# 3x3
s21 = AuxConv(
C.Convolution2D(input_channel, 384, 1, use_cudnn=use_cudnn))
s22 = Inception(
AuxConv(
C.Convolution2D(384,
384, (1, 3),
pad=(0, 1),
use_cudnn=use_cudnn)),
AuxConv(
C.Convolution2D(384,
384, (3, 1),
pad=(1, 0),
use_cudnn=use_cudnn)))
s2 = Sequential(s21, s22)
# double 3x3
s31 = AuxConv(
C.Convolution2D(input_channel, 448, 1, use_cudnn=use_cudnn))
s32 = AuxConv(
C.Convolution2D(448, 384, 3, pad=1, use_cudnn=use_cudnn))
s331 = AuxConv(
C.Convolution2D(384,
384, (1, 3),
pad=(0, 1),
use_cudnn=use_cudnn))
s332 = AuxConv(
C.Convolution2D(384,
384, (3, 1),
pad=(1, 0),
use_cudnn=use_cudnn))
s33 = Inception(s331, s332)
s3 = Sequential(s31, s32, s33)
# pool
s4 = AuxConv(C.Convolution2D(input_channel,
192,
3,
pad=1,
use_cudnn=use_cudnn),
pool=A.AveragePooling2D(3, 1, 1, use_cudnn=use_cudnn))
return Inception(s1, s2, s3, s4)
def __init__(self, in_channels, out1, proj3, out3, proj33, out33,
pooltype, proj_pool=None, stride=1):
if out1 > 0:
assert stride == 1
assert proj_pool is not None
self.f = function_set.FunctionSet(
proj3=convolution_2d.Convolution2D(
in_channels, proj3, 1, nobias=True),
conv3=convolution_2d.Convolution2D(
proj3, out3, 3, pad=1, stride=stride, nobias=True),
proj33=convolution_2d.Convolution2D(
in_channels, proj33, 1, nobias=True),
conv33a=convolution_2d.Convolution2D(
proj33, out33, 3, pad=1, nobias=True),
conv33b=convolution_2d.Convolution2D(
out33, out33, 3, pad=1, stride=stride, nobias=True),
proj3n=batch_normalization.BatchNormalization(proj3),
conv3n=batch_normalization.BatchNormalization(out3),
proj33n=batch_normalization.BatchNormalization(proj33),
conv33an=batch_normalization.BatchNormalization(out33),
conv33bn=batch_normalization.BatchNormalization(out33),
)
if out1 > 0:
self.f.conv1 = convolution_2d.Convolution2D(
in_channels, out1, 1, stride=stride, nobias=True)
self.f.conv1n = batch_normalization.BatchNormalization(out1)
self.out1 = out1
if proj_pool is not None:
self.f.poolp = convolution_2d.Convolution2D(
in_channels, proj_pool, 1, nobias=True)
self.f.poolpn = batch_normalization.BatchNormalization(proj_pool)
self.proj_pool = proj_pool
if pooltype == 'max':
self.f.pool = max_pooling_2d.MaxPooling2D(3, stride=stride, pad=1)
elif pooltype == 'avg':
self.f.pool = average_pooling_2d.AveragePooling2D(
3, stride=stride, pad=1)
else:
raise NotImplementedError()
def inception_35(ich, pch, name):
# 1x1
s1 = ConvBN(ich, 64, 1, 1, 0, name['1x1'][0])
# 5x5
s21 = ConvBN(ich, 48, 1, 1, 0, name['5x5'][0])
s22 = ConvBN(48, 64, 5, 1, 2, name['5x5'][1])
s2 = Sequential(s21, s22)
# double 3x3
s31 = ConvBN(ich, 64, 1, 1, 0, name['3x3'][0])
s32 = ConvBN(64, 96, 3, 1, 1, name['3x3'][1])
s33 = ConvBN(96, 96, 3, 1, 1, name['3x3'][2])
s3 = Sequential(s31, s32, s33)
# pool
s4 = ConvBN(ich,
pch,
1,
1,
0,
name['pool'][1],
pool=A.AveragePooling2D(3, 1, 1))
return Inception(s1, s2, s3, s4)
def __init__(self, use_cudnn=True):
convolution = link.ChainList(
AuxConv(C.Convolution2D(3, 32, 3, 2, use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(32, 32, 3, use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(32, 64, 3, 1, 1, use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(64, 80, 3, 1, 1, use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(80, 192, 3, use_cudnn=use_cudnn)))
def inception_0(input_channel, pool_channel):
# 1x1
s1 = AuxConv(
C.Convolution2D(input_channel, 64, 1, use_cudnn=use_cudnn))
# 5x5
s21 = AuxConv(
C.Convolution2D(input_channel, 48, 1, use_cudnn=use_cudnn))
s22 = AuxConv(
C.Convolution2D(48, 64, 5, pad=2, use_cudnn=use_cudnn))
s2 = Sequential(s21, s22)
# double 3x3
s31 = AuxConv(
C.Convolution2D(input_channel, 64, 1, use_cudnn=use_cudnn))
s32 = AuxConv(
C.Convolution2D(64, 96, 3, pad=1, use_cudnn=use_cudnn))
s33 = AuxConv(
C.Convolution2D(96, 96, 3, pad=1, use_cudnn=use_cudnn))
s3 = Sequential(s31, s32, s33)
# pool
s4 = AuxConv(C.Convolution2D(input_channel,
pool_channel,
3,
pad=1,
use_cudnn=use_cudnn),
pool=M.MaxPooling2D(3, 1, 1, use_cudnn=use_cudnn))
return Inception(s1, s2, s3, s4)
inception0 = Sequential(*[
inception_0(input_channel, pool_channel) for input_channel,
pool_channel in zip([192, 256, 288], [32, 64, 64])
])
grid_reduction0 = Inception(
# strided 3x3
AuxConv(C.Convolution2D(288, 384, 3, 2, use_cudnn=use_cudnn)),
# double 3x3
Sequential(
AuxConv(C.Convolution2D(288, 64, 1, use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(64, 96, 3, pad=1,
use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(96, 96, 3, 2, use_cudnn=use_cudnn))),
# pool
pool=M.MaxPooling2D(3, 2))
def inception_1(hidden_channel):
# 1x1
s1 = AuxConv(C.Convolution2D(768, 192, 1, use_cudnn=use_cudnn))
# 7x7
s21 = AuxConv(
C.Convolution2D(768, hidden_channel, 1, use_cudnn=use_cudnn))
s22 = AuxConv(
C.Convolution2D(hidden_channel,
hidden_channel, (1, 7),
pad=(0, 3),
use_cudnn=use_cudnn))
s23 = AuxConv(
C.Convolution2D(hidden_channel,
192, (7, 1),
pad=(3, 0),
use_cudnn=use_cudnn))
s2 = Sequential(s21, s22, s23)
# double 7x7
s31 = AuxConv(
C.Convolution2D(768, hidden_channel, 1, use_cudnn=use_cudnn))
s32 = AuxConv(
C.Convolution2D(hidden_channel,
hidden_channel, (1, 7),
pad=(0, 3),
use_cudnn=use_cudnn))
s33 = AuxConv(
C.Convolution2D(hidden_channel,
hidden_channel, (7, 1),
pad=(3, 0),
use_cudnn=use_cudnn))
s34 = AuxConv(
C.Convolution2D(hidden_channel,
hidden_channel, (1, 7),
pad=(0, 3),
use_cudnn=use_cudnn))
s35 = AuxConv(
C.Convolution2D(hidden_channel,
192, (7, 1),
pad=(3, 0),
use_cudnn=use_cudnn))
s3 = Sequential(s31, s32, s33, s34, s35)
# pool
s4 = AuxConv(C.Convolution2D(768,
192,
3,
pad=1,
use_cudnn=use_cudnn),
pool=A.AveragePooling2D(3, 1, 1, use_cudnn=use_cudnn))
return Inception(s1, s2, s3, s4)
inception1 = Sequential(
*[inception_1(c) for c in [128, 160, 160, 192]])
grid_reduction1 = Inception(
# strided 3x3
Sequential(
AuxConv(C.Convolution2D(768, 192, 1, use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(192, 320, 3, 2, use_cudnn=use_cudnn))),
# 7x7 and 3x3
Sequential(
AuxConv(C.Convolution2D(768, 192, 1, use_cudnn=use_cudnn)),
AuxConv(
C.Convolution2D(192,
192, (1, 7),
pad=(0, 3),
use_cudnn=use_cudnn)),
AuxConv(
C.Convolution2D(192,
192, (7, 1),
pad=(3, 0),
use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(192, 192, 3, 2, use_cudnn=use_cudnn))),
# pool
pool=M.MaxPooling2D(3, 2, use_cudnn=use_cudnn))
def inception_2(input_channel):
# 1x1
s1 = AuxConv(
C.Convolution2D(input_channel, 320, 1, use_cudnn=use_cudnn))
# 3x3
s21 = AuxConv(
C.Convolution2D(input_channel, 384, 1, use_cudnn=use_cudnn))
s22 = Inception(
AuxConv(
C.Convolution2D(384,
384, (1, 3),
pad=(0, 1),
use_cudnn=use_cudnn)),
AuxConv(
C.Convolution2D(384,
384, (3, 1),
pad=(1, 0),
use_cudnn=use_cudnn)))
s2 = Sequential(s21, s22)
# double 3x3
s31 = AuxConv(
C.Convolution2D(input_channel, 448, 1, use_cudnn=use_cudnn))
s32 = AuxConv(
C.Convolution2D(448, 384, 3, pad=1, use_cudnn=use_cudnn))
s331 = AuxConv(
C.Convolution2D(384,
384, (1, 3),
pad=(0, 1),
use_cudnn=use_cudnn))
s332 = AuxConv(
C.Convolution2D(384,
384, (3, 1),
pad=(1, 0),
use_cudnn=use_cudnn))
s33 = Inception(s331, s332)
s3 = Sequential(s31, s32, s33)
# pool
s4 = AuxConv(C.Convolution2D(input_channel,
192,
3,
pad=1,
use_cudnn=use_cudnn),
pool=A.AveragePooling2D(3, 1, 1, use_cudnn=use_cudnn))
return Inception(s1, s2, s3, s4)
inception2 = Sequential(
*[inception_2(input_channel) for input_channel in [1280, 2048]])
auxiliary_convolution = Sequential(
AuxConv(C.Convolution2D(768, 128, 1, use_cudnn=use_cudnn),
pool=A.AveragePooling2D(5, 3, use_cudnn=use_cudnn)),
AuxConv(C.Convolution2D(128, 768, 5, use_cudnn=use_cudnn)))
super(InceptionV3, self).__init__(
convolution=convolution,
inception=link.ChainList(inception0, inception1, inception2),
grid_reduction=link.ChainList(grid_reduction0, grid_reduction1),
auxiliary_convolution=auxiliary_convolution,
auxiliary_linear=linear.Linear(768, 1000),
linear=linear.Linear(2048, 1000))