def __init__(self,
ch0,
ch1,
bn=True,
sample='down',
activation=F.relu,
dropout=False) -> None:
super().__init__()
self.bn = bn
self.activation = activation
self.dropout = dropout
w = chainer.initializers.Normal(0.02)
with self.init_scope():
if sample == 'down':
self.c = L.Convolution2D(ch0, ch1, 4, 2, 1, initialW=w)
elif sample == 'up':
self.c = L.Deconvolution2D(ch0, ch1, 4, 2, 1, initialW=w)
else:
self.c = L.Convolution2D(ch0, ch1, 1, 1, 0, initialW=w)
if bn:
self.batchnorm = L.BatchNormalization(ch1)
def setUp(self):
args, kwargs = self.deconv_args
kwargs['nobias'] = self.nobias
self.link = L.Deconvolution2D(*args, **kwargs)
if not self.nobias:
self.link.b.data[...] = numpy.random.uniform(
-1, 1, self.link.b.data.shape).astype(numpy.float32)
out_channels = self.link.out_channels
ksize = self.link.ksize
stride = self.link.stride[0]
pad = self.link.pad[0]
N = 2
h, w = 3, 2
kh, kw = _pair(ksize)
out_h = conv.get_deconv_outsize(h, kh, stride, pad)
out_w = conv.get_deconv_outsize(w, kw, stride, pad)
self.gy = numpy.random.uniform(
-1, 1, (N, out_channels, out_h, out_w)).astype(numpy.float32)
self.x = numpy.random.uniform(-1, 1,
(N, 3, h, w)).astype(numpy.float32)
self.link(chainer.Variable(self.x))
self.link.cleargrads()
def __init__(self, n_class=21):
self.n_class = n_class
kwargs = {
'initialW': chainer.initializers.Zero(),
'initial_bias': chainer.initializers.Zero(),
}
super(FCN32s, self).__init__()
with self.init_scope():
self.conv1_1 = L.Convolution2D(3, 64, 3, 1, 100, **kwargs)
self.conv1_2 = L.Convolution2D(64, 64, 3, 1, 1, **kwargs)
self.conv2_1 = L.Convolution2D(64, 128, 3, 1, 1, **kwargs)
self.conv2_2 = L.Convolution2D(128, 128, 3, 1, 1, **kwargs)
self.conv3_1 = L.Convolution2D(128, 256, 3, 1, 1, **kwargs)
self.conv3_2 = L.Convolution2D(256, 256, 3, 1, 1, **kwargs)
self.conv3_3 = L.Convolution2D(256, 256, 3, 1, 1, **kwargs)
self.conv4_1 = L.Convolution2D(256, 512, 3, 1, 1, **kwargs)
self.conv4_2 = L.Convolution2D(512, 512, 3, 1, 1, **kwargs)
self.conv4_3 = L.Convolution2D(512, 512, 3, 1, 1, **kwargs)
self.conv5_1 = L.Convolution2D(512, 512, 3, 1, 1, **kwargs)
self.conv5_2 = L.Convolution2D(512, 512, 3, 1, 1, **kwargs)
self.conv5_3 = L.Convolution2D(512, 512, 3, 1, 1, **kwargs)
self.fc6 = L.Convolution2D(512, 4096, 7, 1, 0, **kwargs)
self.fc7 = L.Convolution2D(4096, 4096, 1, 1, 0, **kwargs)
self.score_fr = L.Convolution2D(4096, n_class, 1, 1, 0, **kwargs)
self.upscore = L.Deconvolution2D(
n_class,
n_class,
64,
32,
0,
nobias=True,
initialW=initializers.UpsamplingDeconvWeight())
def __init__(self, in_ch, out_ch, sample):
super().__init__()
with self.init_scope():
self.bn = L.BatchNormalization(in_ch)
if sample == 'down':
self.c = L.Convolution2D(in_ch,
out_ch,
ksize=4,
stride=2,
pad=1)
elif sample == 'up':
self.c = L.Deconvolution2D(in_ch,
out_ch,
ksize=4,
stride=2,
pad=1)
else:
self.c = L.Convolution2D(in_ch,
out_ch,
ksize=1,
stride=1,
pad=0)
def __init__(self, n_first, n_latent):
super(Decoder, self).__init__()
self.n_first = n_first
with self.init_scope():
self.d_l0 = L.Linear(n_latent, n_first * 16 * 4)
self.d_dc0 = L.Deconvolution2D(in_channels=n_first * 16, out_channels=n_first * 16, ksize=4, stride=2, pad=1)
self.d_dc1 = L.Deconvolution2D(in_channels=n_first * 16, out_channels=n_first * 8, ksize=4, stride=2, pad=1)
self.d_dc2 = L.Deconvolution2D(in_channels=n_first * 8, out_channels=n_first * 4, ksize=4, stride=2, pad=1)
self.d_dc3 = L.Deconvolution2D(in_channels=n_first * 4, out_channels=n_first * 2, ksize=4, stride=2, pad=1)
self.d_dc4 = L.Deconvolution2D(in_channels=n_first * 2, out_channels=n_first, ksize=4, stride=2, pad=1)
self.d_dc5 = L.Deconvolution2D(in_channels=n_first, out_channels=3, ksize=4, stride=2, pad=1)
self.d_bn0 = L.BatchNormalization(n_first * 16, use_gamma=False)
self.d_bn1 = L.BatchNormalization(n_first * 8, use_gamma=False)
self.d_bn2 = L.BatchNormalization(n_first * 4, use_gamma=False)
self.d_bn3 = L.BatchNormalization(n_first * 2, use_gamma=False)
self.d_bn4 = L.BatchNormalization(n_first, use_gamma=False)
self.d_bn5 = L.BatchNormalization(3, use_gamma=False)
def __init__(self):
"""
レイヤー定義
Parameter
---------
chs : int
residualblockのチャンネル数
layres : int
residualblockの数
"""
super(Generator, self).__init__()
he_init = chainer.initializers.HeNormal()
with self.init_scope():
self.e = L.Convolution2D(1025, chs, (4, 1), stride=(4, 1), initialW=he_init).add_hook(spn())
self.c11 = L.Convolution2D(chs, chs, (5, 1), initialW=he_init).add_hook(spn())
self.c12 = L.Convolution2D(chs, chs, (5, 1), initialW=he_init).add_hook(spn())
self.c21 = L.Convolution2D(chs, chs, (5, 1), initialW=he_init).add_hook(spn())
self.c22 = L.Convolution2D(chs, chs, (5, 1), initialW=he_init).add_hook(spn())
self.c31 = L.Convolution2D(chs, chs, (5, 1), initialW=he_init).add_hook(spn())
self.c32 = L.Convolution2D(chs, chs, (5, 1), initialW=he_init).add_hook(spn())
self.c41 = L.Convolution2D(chs, chs, (5, 1), initialW=he_init).add_hook(spn())
self.c42 = L.Convolution2D(chs, chs, (5, 1), initialW=he_init).add_hook(spn())
self.d = L.Deconvolution2D(chs, 1025, (4, 1), stride=(4, 1), initialW=he_init).add_hook(spn())
def __init__(self, n_class=4):
self.n_class = n_class
super(self.__class__, self).__init__(
conv1_1=L.Convolution2D(3, 64, 3, stride=1, pad=100),
conv1_2=L.Convolution2D(64, 64, 3, stride=1, pad=1),
conv2_1=L.Convolution2D(64, 128, 3, stride=1, pad=1),
conv2_2=L.Convolution2D(128, 128, 3, stride=1, pad=1),
conv3_1=L.Convolution2D(128, 256, 3, stride=1, pad=1),
conv3_2=L.Convolution2D(256, 256, 3, stride=1, pad=1),
conv3_3=L.Convolution2D(256, 256, 3, stride=1, pad=1),
conv5_1=L.Convolution2D(256, 256, 3, stride=1, pad=1),
conv5_2=L.Convolution2D(256, 256, 3, stride=1, pad=1),
conv5_3=L.Convolution2D(256, 256, 3, stride=1, pad=1),
fc6=L.Convolution2D(256, 512, 7, stride=1, pad=0),
fc7=L.Convolution2D(512, 512, 1, stride=1, pad=0),
score_fr=L.Convolution2D(512, self.n_class, 1, stride=1, pad=0),
upscore=L.Deconvolution2D(self.n_class,
self.n_class,
64,
stride=32,
pad=0),
)
self.train = False
def __init__(self, in_channels, out_channels, LoopTimes=5):
super(Block, self).__init__()
with self.init_scope():
"""
ffconv : feedforword Convolution.
fbconv : feedback Convolution (deconvolution).
bpconv : bypass Convolution (1x1 convolution).
update_late : A learnable and non-negative parameter.
"""
self.bn = L.BatchNormalization(in_channels)
self.ffconv = L.Convolution2D(in_channels,
out_channels,
ksize=3,
pad=1)
self.fbconv = L.Deconvolution2D(out_channels,
in_channels,
ksize=3,
pad=1)
self.bpconv = L.Convolution2D(in_channels, out_channels, ksize=1)
self.update_rate = L.Scale(axis=0, W_shape=(1, ))
self.LoopTimes = LoopTimes
self.out_channels = out_channels
def __init__(self,
input_ch,
out_ch,
ksize=None,
stride=1,
pad=0,
nobias=False,
initialW=None,
initial_bias=None,
outsize=None):
super().__init__()
with self.init_scope():
self.c = L.Deconvolution2D(input_ch,
out_ch,
ksize=ksize,
stride=stride,
pad=pad,
nobias=nobias,
initialW=initialW,
initial_bias=initial_bias,
outsize=outsize)
self.c.add_hook(SpectralNormalization())
def __init__(self,
n_class,
roi_size,
spatial_scale,
vgg_initialW=None,
loc_initialW=None,
score_initialW=None,
mask_initialW=None,
pooling_func=functions.roi_align_2d):
# n_class includes the background
super(VGG16RoIHead, self).__init__()
with self.init_scope():
# self.fc6 = L.Linear(25088, 4096, initialW=vgg_initialW)
self.fc6 = L.Linear(512 * roi_size * roi_size,
4096,
initialW=vgg_initialW)
self.fc7 = L.Linear(4096, 4096, initialW=vgg_initialW)
self.cls_loc = L.Linear(4096, n_class * 4, initialW=loc_initialW)
self.score = L.Linear(4096, n_class, initialW=score_initialW)
# 7 x 7 x 512 -> 14 x 14 x 256
self.deconv6 = L.Deconvolution2D(512,
256,
2,
stride=2,
initialW=mask_initialW)
# 14 x 14 x 256 -> 14 x 14 x 20
n_fg_class = n_class - 1
self.mask = L.Convolution2D(256,
n_fg_class,
1,
initialW=mask_initialW)
self.n_class = n_class
self.roi_size = roi_size
self.spatial_scale = spatial_scale
self._pooling_func = pooling_func
def __init__(self, noise=128, dim=64):
super(Generator, self).__init__()
self.dim = dim
with self.init_scope():
self.fc = L.Linear(None, dim * 8 * 4 * 4)
self.res1 = Block(6, 8 * dim, 8 * dim, norm=L.BatchNormalization)
self.res1up = Block(1,
8 * dim,
4 * dim,
resample='up',
norm=L.BatchNormalization)
self.res2 = Block(6, 4 * dim, 4 * dim, norm=L.BatchNormalization)
self.res2up = Block(1,
4 * dim,
2 * dim,
resample='up',
norm=L.BatchNormalization)
self.res3 = Block(6, 2 * dim, 2 * dim, norm=L.BatchNormalization)
self.res3up = Block(1,
2 * dim,
1 * dim,
resample='up',
norm=L.BatchNormalization)
self.res4 = Block(6, 1 * dim, 1 * dim, norm=L.BatchNormalization)
self.res4up = Block(1,
1 * dim,
dim // 2,
resample='up',
norm=L.BatchNormalization)
self.res5 = Block(5, dim // 2, dim // 2, norm=L.BatchNormalization)
self.conv = L.Deconvolution2D(None, 3, 3, pad=1)
self.noise = noise
def __init__(self, channels_chrz, channels_u):
super().__init__(lstm_tanh=L.Convolution2D(None,
channels_chrz,
ksize=5,
stride=1,
pad=2),
lstm_i=L.Convolution2D(None,
channels_chrz,
ksize=5,
stride=1,
pad=2),
lstm_f=L.Convolution2D(None,
channels_chrz,
ksize=5,
stride=1,
pad=2),
lstm_o=L.Convolution2D(None,
channels_chrz,
ksize=5,
stride=1,
pad=2),
mean_z=L.Convolution2D(None,
channels_chrz,
ksize=5,
stride=1,
pad=2),
mean_x=L.Convolution2D(None,
3,
ksize=1,
stride=1,
pad=0),
deconv_h=L.Deconvolution2D(None,
channels_u,
ksize=4,
stride=4,
pad=0))
def setUp(self):
out_channels = 2
ksize = 3
stride = 2
pad = 1
in_channels = None
self.link = L.Deconvolution2D(
in_channels, out_channels, ksize,
stride=stride, pad=pad, nobias=self.nobias)
if not self.nobias:
self.link.b.data[...] = numpy.random.uniform(
-1, 1, self.link.b.data.shape).astype(numpy.float32)
N = 2
h, w = 3, 2
kh, kw = _pair(ksize)
out_h = conv.get_deconv_outsize(h, kh, stride, pad)
out_w = conv.get_deconv_outsize(w, kw, stride, pad)
self.gy = numpy.random.uniform(
-1, 1, (N, out_channels, out_h, out_w)).astype(numpy.float32)
self.x = numpy.random.uniform(
-1, 1, (N, 3, h, w)).astype(numpy.float32)
self.link(chainer.Variable(self.x))
self.link.cleargrads()
def __init__(self, latent_size, wscale=0.02):
w = chainer.initializers.Normal(wscale)
super(Generator, self).__init__()
self.latent_size = latent_size
with self.init_scope():
self.fc1 = BatchLinear(latent_size, 4 * 4 * 128, initialW=w)
self.deconv2 = BatchDeconv2D(128,
64,
4,
pad=1,
stride=2,
initialW=w)
self.deconv3 = BatchDeconv2D(64,
32,
4,
pad=1,
stride=2,
initialW=w)
self.deconv4 = L.Deconvolution2D(32,
1,
4,
pad=1,
stride=2,
initialW=w)
def __init__(self,
ch0,
ch1,
bn=True,
sample='down',
activation=F.leaky_relu,
add_noise=False,
sigma=0.2):
self.bn = bn
self.activation = activation
self.add_noise = add_noise
self.sigma = sigma
self.iteration = 0
layers = {}
w = chainer.initializers.Normal(0.02)
if sample == 'down':
layers['c'] = L.Convolution2D(ch0, ch1, 4, 2, 1, initialW=w)
elif sample == 'up':
layers['c'] = L.Deconvolution2D(ch0, ch1, 4, 2, 1, initialW=w)
elif sample == 'same':
layers['c'] = L.Convolution2D(ch0, ch1, 3, 1, 1, initialW=w)
if bn:
layers['batchnorm'] = L.BatchNormalization(ch1)
super(CBR, self).__init__(**layers)
def __init__(self, n_units, dim_z, directory=None):
super(TdVae, self).__init__(
conv1=L.Convolution2D(1, 20, 9),
lin1=L.Linear(20 * 20 * 20, 400),
encoder=L.LSTM(400, out_size=n_units),
bn=chainer.links.BatchNormalization(n_units),
p_b_2_lin=L.Linear(n_units),
p_b_2_m=L.Linear(dim_z),
p_b_2_v=L.Linear(dim_z),
q_I_lin=L.Linear(n_units),
q_I_m=L.Linear(dim_z),
q_I_v=L.Linear(dim_z),
p_b_1_lin=L.Linear(n_units),
p_b_1_m=L.Linear(dim_z),
p_b_1_v=L.Linear(dim_z),
p_p_2_lin=L.Linear(n_units),
p_p_2_m=L.Linear(dim_z),
p_p_2_v=L.Linear(dim_z),
p_d_1=L.Linear(20 * 20 * 20),
p_d_2=L.Linear(784),
p_d_conv=L.Deconvolution2D(20, 1, 9),
)
self.it = 0
self.directory = directory
def setUp(self):
self.link = L.Deconvolution2D(self.in_channels,
self.out_channels,
self.ksize,
stride=self.stride,
pad=self.pad,
nobias=self.nobias)
self.link.W.data[...] = numpy.random.uniform(
-1, 1, self.link.W.data.shape).astype(numpy.float32)
if not self.nobias:
self.link.b.data[...] = numpy.random.uniform(
-1, 1, self.link.b.data.shape).astype(numpy.float32)
self.link.zerograds()
N = 2
h, w = 3, 2
kh, kw = _pair(self.ksize)
out_h = conv.get_deconv_outsize(h, kh, self.stride, self.pad)
out_w = conv.get_deconv_outsize(w, kw, self.stride, self.pad)
self.gy = numpy.random.uniform(
-1, 1, (N, self.out_channels, out_h, out_w)).astype(numpy.float32)
self.x = numpy.random.uniform(
-1, 1, (N, self.in_channels, h, w)).astype(numpy.float32)
def __init__(self, activation=F.relu):
super(DispNet, self).__init__()
self.activation = activation
with self.init_scope():
self.c1 = L.Convolution2D(None, 32, ksize=7, stride=2, pad=3)
self.c1b = L.Convolution2D(None, 32, ksize=7, stride=1, pad=3)
self.c2 = L.Convolution2D(None, 64, ksize=5, stride=2, pad=2)
self.c2b = L.Convolution2D(None, 64, ksize=5, stride=1, pad=2)
self.c3 = L.Convolution2D(None, 128, ksize=3, stride=2, pad=1)
self.c3b = L.Convolution2D(None, 128, ksize=3, stride=1, pad=1)
self.c4 = L.Convolution2D(None, 256, ksize=3, stride=2, pad=1)
self.c4b = L.Convolution2D(None, 256, ksize=3, stride=1, pad=1)
self.c5 = L.Convolution2D(None, 512, ksize=3, stride=2, pad=1)
self.c5b = L.Convolution2D(None, 512, ksize=3, stride=1, pad=1)
self.c6 = L.Convolution2D(None, 512, ksize=3, stride=2, pad=1)
self.c6b = L.Convolution2D(None, 512, ksize=3, stride=1, pad=1)
self.c7 = L.Convolution2D(None, 512, ksize=3, stride=2, pad=1)
self.c7b = L.Convolution2D(None, 512, ksize=3, stride=1, pad=1)
self.dc7 = L.Deconvolution2D(None, 512, ksize=4, stride=2, pad=1)
self.idc7 = L.Convolution2D(None, 512, ksize=3, stride=1, pad=1)
self.dc6 = L.Deconvolution2D(None, 512, ksize=4, stride=2, pad=1)
self.idc6 = L.Convolution2D(None, 512, ksize=3, stride=1, pad=1)
self.dc5 = L.Deconvolution2D(None, 256, ksize=4, stride=2, pad=1)
self.idc5 = L.Convolution2D(None, 256, ksize=3, stride=1, pad=1)
self.dc4 = L.Deconvolution2D(None, 128, ksize=4, stride=2, pad=1)
self.idc4 = L.Convolution2D(None, 128, ksize=3, stride=1, pad=1)
self.dispout4 = L.Convolution2D(None, 1, ksize=3, stride=1, pad=1)
self.dc3 = L.Deconvolution2D(None, 64, ksize=4, stride=2, pad=1)
self.idc3 = L.Convolution2D(None, 64, ksize=3, stride=1, pad=1)
self.dispout3 = L.Convolution2D(None, 1, ksize=3, stride=1, pad=1)
self.dc2 = L.Deconvolution2D(None, 32, ksize=4, stride=2, pad=1)
self.idc2 = L.Convolution2D(None, 32, ksize=3, stride=1, pad=1)
self.dispout2 = L.Convolution2D(None, 1, ksize=3, stride=1, pad=1)
self.dc1 = L.Deconvolution2D(None, 16, ksize=4, stride=2, pad=1)
self.idc1 = L.Convolution2D(None, 16, ksize=3, stride=1, pad=1)
self.dispout1 = L.Convolution2D(None, 1, ksize=3, stride=1, pad=1)
def __init__(self):
super(Generator, self).__init__(
dc0z=L.Deconvolution2D(nz,
512,
args.final_filter_size,
stride=args.stride,
wscale=0.02 * math.sqrt(nz)),
dc0s=L.Deconvolution2D(n_signal,
512,
args.final_filter_size,
stride=args.stride,
wscale=0.02 * math.sqrt(n_signal)),
dc1=L.Deconvolution2D(512,
256,
4,
stride=2,
pad=1,
wscale=0.02 * math.sqrt(4 * 4 * 512)),
dc2=L.Deconvolution2D(256,
128,
4,
stride=2,
pad=1,
wscale=0.02 * math.sqrt(4 * 4 * 256)),
dc3=L.Deconvolution2D(128,
64,
4,
stride=2,
pad=1,
wscale=0.02 * math.sqrt(4 * 4 * 128)),
dc4=L.Deconvolution2D(64,
1,
4,
stride=2,
pad=1,
wscale=0.02 * math.sqrt(4 * 4 * 64)),
bn0=L.BatchNormalization(512),
bn1=L.BatchNormalization(256),
bn2=L.BatchNormalization(128),
bn3=L.BatchNormalization(64),
)
def __init__(self):
super(CAE, self).__init__()
with self.init_scope():
self.conv1_1 = chainercv.links.Conv2DBNActiv(None,
32,
3,
stride=1,
pad=1)
self.conv1_2 = chainercv.links.Conv2DBNActiv(None,
32,
3,
stride=1,
pad=1)
self.conv2_1 = chainercv.links.Conv2DBNActiv(None,
32,
3,
stride=1,
pad=1)
self.conv2_2 = chainercv.links.Conv2DBNActiv(None,
32,
3,
stride=1,
pad=1)
self.conv3_1 = chainercv.links.Conv2DBNActiv(None,
32,
3,
stride=1,
pad=1)
self.conv3_2 = chainercv.links.Conv2DBNActiv(None,
32,
3,
stride=1,
pad=1)
self.deconv1_1 = L.Deconvolution2D(None, 32, 3, stride=1, pad=1)
self.deconv1_2 = L.Deconvolution2D(None, 32, 3, stride=1, pad=1)
self.deconv2_1 = L.Deconvolution2D(None, 32, 3, stride=1, pad=1)
self.deconv2_2 = L.Deconvolution2D(None, 32, 3, stride=1, pad=1)
self.deconv3_1 = L.Deconvolution2D(None, 32, 3, stride=1, pad=1)
self.deconv3_2 = L.Deconvolution2D(None, 1, 3, stride=1, pad=1)
def __init__(self, act=F.relu):
super(Decoder, self).__init__(
linear0=L.Linear(10, 32),
linear1=L.Linear(32, 64 * 7 * 7),
deconv0=L.Deconvolution2D(64, 64, 2, stride=2, pad=0),
deconv1=L.Deconvolution2D(64, 64, 3, stride=1, pad=1),
deconv2=L.Deconvolution2D(64, 64, 3, stride=1, pad=1),
deconv3=L.Deconvolution2D(64, 32, 2, stride=2, pad=0),
deconv4=L.Deconvolution2D(32, 32, 3, stride=1, pad=1),
deconv5=L.Deconvolution2D(32, 1, 3, stride=1, pad=1),
bn_linear0=L.BatchNormalization(32, decay=0.9, use_cudnn=True),
bn_linear1=L.BatchNormalization(64 * 7 * 7,
decay=0.9,
use_cudnn=True),
bn_deconv0=L.BatchNormalization(64, decay=0.9, use_cudnn=True),
bn_deconv1=L.BatchNormalization(64, decay=0.9, use_cudnn=True),
bn_deconv2=L.BatchNormalization(64, decay=0.9, use_cudnn=True),
bn_deconv3=L.BatchNormalization(32, decay=0.9, use_cudnn=True),
bn_deconv4=L.BatchNormalization(32, decay=0.9, use_cudnn=True),
)
self.act = act
self.hiddens = []
def __init__(self,
feature_map_nc,
output_nc,
tanh_constant,
instance_normalization=True,
w_init=None):
self.tanh_constant = tanh_constant
BN = InstanceNormalization if instance_normalization else L.BatchNormalization
super(ImageTransformer, self).__init__(
c1=L.Convolution2D(None,
feature_map_nc,
ksize=9,
stride=1,
pad=4,
initialW=w_init),
c2=L.Convolution2D(None,
2 * feature_map_nc,
ksize=3,
stride=2,
pad=1,
initialW=w_init),
c3=L.Convolution2D(None,
4 * feature_map_nc,
ksize=3,
stride=2,
pad=1,
initialW=w_init),
r1=ResidualBlock(4 * feature_map_nc,
instance_normalization=instance_normalization,
w_init=w_init),
r2=ResidualBlock(4 * feature_map_nc,
instance_normalization=instance_normalization,
w_init=w_init),
r3=ResidualBlock(4 * feature_map_nc,
instance_normalization=instance_normalization,
w_init=w_init),
r4=ResidualBlock(4 * feature_map_nc,
instance_normalization=instance_normalization,
w_init=w_init),
r5=ResidualBlock(4 * feature_map_nc,
instance_normalization=instance_normalization,
w_init=w_init),
d1=L.Deconvolution2D(None,
2 * feature_map_nc,
ksize=4,
stride=2,
pad=1,
initialW=w_init),
d2=L.Deconvolution2D(None,
feature_map_nc,
ksize=4,
stride=2,
pad=1,
initialW=w_init),
d3=L.Convolution2D(None,
output_nc,
ksize=9,
stride=1,
pad=4,
initialW=w_init),
b1=BN(feature_map_nc),
b2=BN(2 * feature_map_nc),
b3=BN(4 * feature_map_nc),
b4=BN(2 * feature_map_nc),
b5=BN(feature_map_nc))
def __init__(self, imap, omap, k=4, s=2, p=1, act=F.relu):
super(DeconvUnit, self).__init__(
deconv=L.Deconvolution2D(imap, omap, ksize=k, stride=s, pad=p, ),
bn=L.BatchNormalization(omap, decay=0.9, use_cudnn=True),
)
self.act = act
def grouped_deconv():
conv = L.Deconvolution2D(16, 32, ksize=3, pad=1, groups=4)
y = conv(x)
return {'input': x}, {'out': y}
def __init__(self, n_in, n_out, ksize, stride, pad):
super(Deconv, self).__init__()
with self.init_scope():
self.conv = L.Deconvolution2D(n_in, n_out, ksize, stride, pad)
self.bn = L.BatchNormalization(n_out)
def deconvolution():
conv = L.Deconvolution2D(16, 32, ksize=3, pad=1)
y = conv(x)
return {'input': x}, {'out': y}
def __init__(self):
super(SSDNet, self).__init__(
conv1_1=L.Convolution2D(3, 64, ksize=3, stride=1, pad=1),
conv1_2=L.Convolution2D(64, 64, ksize=3, stride=1, pad=1),
conv2_1=L.Convolution2D(64, 128, ksize=3, stride=1, pad=1),
conv2_2=L.Convolution2D(128, 128, ksize=3, stride=1, pad=1),
conv3_1=L.Convolution2D(128, 256, ksize=3, stride=1, pad=1),
conv3_2=L.Convolution2D(256, 256, ksize=3, stride=1, pad=1),
conv3_3=L.Convolution2D(256, 256, ksize=3, stride=1, pad=1),
conv4_1=L.Convolution2D(256, 512, ksize=3, stride=1, pad=1),
conv4_2=L.Convolution2D(512, 512, ksize=3, stride=1, pad=1),
conv4_3=L.Convolution2D(512, 512, ksize=3, stride=1, pad=1),
conv5_1=L.Convolution2D(512, 512, ksize=3, stride=1, pad=1),
conv5_2=L.Convolution2D(512, 512, ksize=3, stride=1, pad=1),
conv5_3=L.Convolution2D(512, 512, ksize=3, stride=1, pad=1),
fc6=L.Convolution2D(512, 1024, ksize=3, stride=1, pad=1),
fc7=L.Convolution2D(1024, 1024, ksize=1, stride=1, pad=0),
conv6_1=L.Convolution2D(1024, 256, ksize=1, stride=1, pad=0),
conv6_2=L.Convolution2D(256, 512, ksize=3, stride=2, pad=1),
bn_conv6=L.BatchNormalization(512),
conv7_1=L.Convolution2D(512, 128, ksize=1, stride=1, pad=0),
conv7_2=L.Convolution2D(128, 256, ksize=3, stride=2, pad=1),
conv8_1=L.Convolution2D(256, 128, ksize=1, stride=1, pad=0),
conv8_2=L.Convolution2D(128, 256, ksize=3, stride=1, pad=0),
conv9_1=L.Convolution2D(256, 128, ksize=1, stride=1, pad=0),
conv9_2=L.Convolution2D(128, 256, ksize=3, stride=1, pad=0),
conv9_2_mbox_loc=L.Convolution2D(256,
common_params.num_of_offset_dims *
common_params.num_boxes[5],
ksize=3,
stride=1,
pad=1),
conv9_2_mbox_cls=L.Convolution2D(256,
common_params.num_of_classes *
common_params.num_boxes[5],
ksize=3,
stride=1,
pad=1),
deconv9_2_1=L.Deconvolution2D(256,
common_params.num_of_classes,
ksize=10,
stride=1,
pad=0), #10×10
deconv9_2_2=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=11,
stride=8,
pad=4), #75×75
deconv9_2_3=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=6,
stride=4,
pad=1), #300×300
deconv1=L.Deconvolution2D(256, 256, ksize=3, stride=1,
pad=0), #3×3
conv8_2_mbox_loc=L.Convolution2D(256,
common_params.num_of_offset_dims *
common_params.num_boxes[4],
ksize=3,
stride=1,
pad=1),
conv8_2_mbox_cls=L.Convolution2D(256,
common_params.num_of_classes *
common_params.num_boxes[4],
ksize=3,
stride=1,
pad=1),
deconv8_2_1=L.Deconvolution2D(256,
common_params.num_of_classes,
ksize=9,
stride=6,
pad=1), #19×19
deconv8_2_2=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=7,
stride=4,
pad=2), #75×75
deconv8_2_3=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=6,
stride=4,
pad=1), #300×300
deconv2=L.Deconvolution2D(256, 256, ksize=3, stride=2,
pad=1), #5×5
conv7_2_mbox_loc=L.Convolution2D(256,
common_params.num_of_offset_dims *
common_params.num_boxes[3],
ksize=3,
stride=1,
pad=1),
conv7_2_mbox_cls=L.Convolution2D(256,
common_params.num_of_classes *
common_params.num_boxes[3],
ksize=3,
stride=1,
pad=1),
deconv7_2_1=L.Deconvolution2D(256,
common_params.num_of_classes,
ksize=7,
stride=4,
pad=2), #19×19
deconv7_2_2=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=7,
stride=4,
pad=2), #75×75
deconv7_2_3=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=6,
stride=4,
pad=1), #300×300
deconv3=L.Deconvolution2D(256, 512, ksize=4, stride=2,
pad=1), #10×10
conv6_2_mbox_loc=L.Convolution2D(512,
common_params.num_of_offset_dims *
common_params.num_boxes[2],
ksize=3,
stride=1,
pad=1),
conv6_2_mbox_cls=L.Convolution2D(512,
common_params.num_of_classes *
common_params.num_boxes[2],
ksize=3,
stride=1,
pad=1),
deconv6_2_1=L.Deconvolution2D(512,
common_params.num_of_classes,
ksize=6,
stride=4,
pad=2), #38×38
deconv6_2_2=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=3,
stride=2,
pad=1), #75×75
deconv6_2_3=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=6,
stride=4,
pad=1), #300×300
deconv4=L.Deconvolution2D(512, 1024, ksize=3, stride=2,
pad=1), #19×19
fc7_mbox_loc=L.Convolution2D(1024,
common_params.num_of_offset_dims *
common_params.num_boxes[1],
ksize=3,
stride=1,
pad=1),
fc7_mbox_cls=L.Convolution2D(1024,
common_params.num_of_classes *
common_params.num_boxes[1],
ksize=3,
stride=1,
pad=1),
deconvfc7_1=L.Deconvolution2D(1024,
common_params.num_of_classes,
ksize=4,
stride=2,
pad=1), #38×38
deconvfc7_2=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=3,
stride=2,
pad=1), #75×75
deconvfc7_3=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=6,
stride=4,
pad=1), #300×300
deconv5=L.Deconvolution2D(1024, 512, ksize=4, stride=2,
pad=1), #38×38
bn4_3=L.BatchNormalization(512),
conv4_3_norm_mbox_loc=L.Convolution2D(
512,
common_params.num_of_offset_dims * common_params.num_boxes[0],
ksize=3,
stride=1,
pad=1),
conv4_3_norm_mbox_cls=L.Convolution2D(
512,
common_params.num_of_classes * common_params.num_boxes[0],
ksize=3,
stride=1,
pad=1),
deconv4_3_1=L.Deconvolution2D(512,
common_params.num_of_classes,
ksize=3,
stride=2,
pad=1), #75×75
deconv4_3_2=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=4,
stride=2,
pad=1), #150×150
deconv4_3_3=L.Deconvolution2D(common_params.num_of_classes,
common_params.num_of_classes,
ksize=4,
stride=2,
pad=1), #300×300
segconv=L.Convolution2D(common_params.num_of_classes * 6,
common_params.num_of_classes,
ksize=1,
stride=1,
pad=0))
self.train = False
def __init__(self, d=256):
super(Decoder, self).__init__()
with self.init_scope():
self.dconv1 = L.Deconvolution2D(d, d, 4, 2, 1)
self.dconv2 = L.Deconvolution2D(d, 1, 4, 2, 1)
def __init__(self, train=False):
self.train = train
base = 64
super(Mynet, self).__init__()
with self.init_scope():
self.enc1 = chainer.Sequential()
for i in range(2):
self.enc1.append(
L.Convolution2D(None,
base,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.enc1.append(F.relu)
self.enc1.append(L.BatchNormalization(base))
self.enc2 = chainer.Sequential()
for i in range(2):
self.enc2.append(
L.Convolution2D(None,
base * 2,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.enc2.append(F.relu)
self.enc2.append(L.BatchNormalization(base * 2))
self.enc3 = chainer.Sequential()
for i in range(2):
self.enc3.append(
L.Convolution2D(None,
base * 4,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.enc3.append(F.relu)
self.enc3.append(L.BatchNormalization(base * 4))
self.enc4 = chainer.Sequential()
for i in range(2):
self.enc4.append(
L.Convolution2D(None,
base * 8,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.enc4.append(F.relu)
self.enc4.append(L.BatchNormalization(base * 8))
self.enc5 = chainer.Sequential()
for i in range(2):
self.enc5.append(
L.Convolution2D(None,
base * 16,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.enc5.append(F.relu)
self.enc5.append(L.BatchNormalization(base * 16))
self.upsample4 = chainer.Sequential()
self.upsample4.append(
L.Deconvolution2D(None, base * 8, ksize=2, stride=2))
self.upsample4.append(F.relu)
self.upsample4.append(L.BatchNormalization(base * 8))
self.dec4 = chainer.Sequential()
for i in range(2):
self.dec4.append(
L.Convolution2D(None,
base * 8,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.dec4.append(F.relu)
self.dec4.append(L.BatchNormalization(base * 8))
self.upsample3 = chainer.Sequential()
self.upsample3.append(
L.Deconvolution2D(None, base * 4, ksize=2, stride=2))
self.upsample3.append(F.relu)
self.upsample3.append(L.BatchNormalization(base * 4))
self.dec3 = chainer.Sequential()
for i in range(2):
self.dec3.append(
L.Convolution2D(None,
base * 4,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.dec3.append(F.relu)
self.dec3.append(L.BatchNormalization(base * 4))
self.upsample2 = chainer.Sequential()
self.upsample2.append(
L.Deconvolution2D(None, base * 2, ksize=2, stride=2))
self.upsample2.append(F.relu)
self.upsample2.append(L.BatchNormalization(base * 2))
self.dec2 = chainer.Sequential()
for i in range(2):
self.dec2.append(
L.Convolution2D(None,
base * 2,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.dec2.append(F.relu)
self.dec2.append(L.BatchNormalization(base * 2))
self.upsample1 = chainer.Sequential()
self.upsample1.append(
L.Deconvolution2D(None, base, ksize=2, stride=2))
self.upsample1.append(F.relu)
self.upsample1.append(L.BatchNormalization(base))
self.dec1 = chainer.Sequential()
for i in range(2):
self.dec1.append(
L.Convolution2D(None,
base,
ksize=3,
pad=1,
stride=1,
nobias=True))
self.dec1.append(F.relu)
self.dec1.append(L.BatchNormalization(base))
self.out = L.Convolution2D(None,
num_classes + 1,
ksize=1,
pad=0,
stride=1,
nobias=False)
def __init__(self,
l_latent=64,
l_seq=32,
mode='generator',
bn=False,
activate_func=F.leaky_relu,
vertical_ksize=1):
if l_seq % 32 != 0:
raise ValueError('\'l_seq\' must be divisible by 32.')
if not mode in ['discriminator', 'generator']:
raise ValueError(
'\'mode\' must be \'discriminator\' or \'generator\'.')
super(ConvAE, self).__init__()
self.bn = bn
self.mode = mode
self.activate_func = activate_func
w = chainer.initializers.Normal(0.02)
with self.init_scope():
self.conv1 = L.Convolution2D(None,
32,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.conv2 = L.Convolution2D(32,
64,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.conv3 = L.Convolution2D(64,
64,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.conv4 = L.Convolution2D(64,
64,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.conv5 = L.Convolution2D(64,
128,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.enc_l = L.Linear(128 * 34 * l_seq // 32, l_latent, initialW=w)
if self.mode == 'generator':
self.dec_l = L.Linear(l_latent,
128 * 17 * l_seq // 32,
initialW=w)
self.deconv1 = L.Deconvolution2D(
128,
64,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.deconv2 = L.Deconvolution2D(
64,
64,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.deconv3 = L.Deconvolution2D(
64,
64,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.deconv4 = L.Deconvolution2D(
64,
32,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
self.deconv5 = L.Deconvolution2D(
32,
1,
ksize=(4, vertical_ksize),
stride=(2, 1),
pad=(1, (vertical_ksize - 1) // 2),
initialW=w)
if self.bn:
self.enc_bn1 = L.BatchNormalization(32)
self.enc_bn2 = L.BatchNormalization(64)
self.enc_bn3 = L.BatchNormalization(64)
self.enc_bn4 = L.BatchNormalization(64)
self.enc_bn5 = L.BatchNormalization(128)
if self.mode == 'generator':
self.dec_bn1 = L.BatchNormalization(128)
self.dec_bn2 = L.BatchNormalization(64)
self.dec_bn3 = L.BatchNormalization(64)
self.dec_bn4 = L.BatchNormalization(64)
self.dec_bn5 = L.BatchNormalization(32)