def __init__(self,
ch=512,
ch_in=512,
w_ch=512,
upsample=True,
enable_blur=False):
super().__init__()
self.upsample = upsample
self.ch = ch
self.ch_in = ch_in
with self.init_scope():
if not upsample:
self.W = chainer.Parameter(shape=(ch_in, 4, 4))
self.W.data[:] = 1 # w_data_tmp
self.b0 = L.Bias(axis=1, shape=(ch, ))
self.b1 = L.Bias(axis=1, shape=(ch, ))
self.n0 = NoiseBlock(ch)
self.n1 = NoiseBlock(ch)
self.s0 = StyleBlock(w_ch, ch)
self.s1 = StyleBlock(w_ch, ch)
self.c0 = EqualizedConv2d(ch_in, ch, 3, 1, 1, nobias=True)
self.c1 = EqualizedConv2d(ch, ch, 3, 1, 1, nobias=True)
self.blur_k = None
self.enable_blur = enable_blur
def __init__(self, channels, filter_width, embeddings, k_top, beta, pool_size):
vocab_size, embed_size = embeddings.shape
feature_size = [
pool_size[0][0] * pool_size[0][1],
pool_size[1][0] * pool_size[1][1] * channels[0],
pool_size[2][0] * pool_size[2][1] * channels[1],
pool_size[3][0] * pool_size[3][1] * channels[1],
]
# initialize functions with parameters
super(BiCNN, self).__init__(
embed=L.EmbedID(
in_size=vocab_size,
out_size=embed_size,
initialW=embeddings,
),
# convolutions in the first block
conv1l=self._create_convolution(
in_channels=1,
out_channels=channels[0],
window_size=filter_width[0],
),
conv1r=self._create_convolution(
in_channels=1,
out_channels=channels[0],
window_size=filter_width[0],
),
bias1l=L.Bias(axis=1, shape=(channels[0], ceil(embed_size / 2))),
bias1r=L.Bias(axis=1, shape=(channels[0], ceil(embed_size / 2))),
# convolutions in the first block
conv2l=self._create_convolution(
in_channels=channels[0],
out_channels=channels[1],
window_size=filter_width[1],
),
conv2r=self._create_convolution(
in_channels=channels[0],
out_channels=channels[1],
window_size=filter_width[1],
),
bias2l=L.Bias(axis=1, shape=(channels[1], ceil(embed_size / 4))),
bias2r=L.Bias(axis=1, shape=(channels[1], ceil(embed_size / 4))),
# output layer
linear=L.Linear(
in_size=sum(feature_size),
out_size=1,
bias=0,
),
)
# retain parameters
self._channels = channels
self._filter_width = filter_width
# self._feature_size = feature_size
self._vocab_size = vocab_size
self._embed_size = embed_size
self._k_top = k_top
self._beta = beta
self._pool_size = pool_size
def __init__(self):
super(TINY_D_36ch, self).__init__(
conv1=L.Convolution2D(36,
128,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn1=L.BatchNormalization(128, use_beta=False),
bias1=L.Bias(shape=(128, )),
# MaxPool(3x3, 2)
conv2=L.Convolution2D(128,
256,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn2=L.BatchNormalization(256, use_beta=False),
bias2=L.Bias(shape=(256, )),
# MaxPool(3x3, 2)
conv3=L.Convolution2D(256,
512,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn3=L.BatchNormalization(512, use_beta=False),
bias3=L.Bias(shape=(512, )),
# MaxPool(3x3, 2)
conv4=L.Convolution2D(512,
1024,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn4=L.BatchNormalization(1024, use_beta=False),
bias4=L.Bias(shape=(1024, )),
conv5=L.Convolution2D(1024,
1024,
ksize=1,
stride=1,
pad=0,
nobias=True),
bn5=L.BatchNormalization(1024, use_beta=False),
bias5=L.Bias(shape=(1024, )),
# AvgPool(3x3, 2)
conv6=L.Convolution2D(1024,
64,
ksize=1,
stride=1,
pad=0,
nobias=True),
conv7=L.Convolution2D(64, 2, ksize=1, stride=1, pad=0),
)
def __init__(self, dim):
super(Denoise, self).__init__(
a0=L.Scale(W_shape=(dim, )),
a1=L.Scale(W_shape=(dim, )),
a2=L.Scale(W_shape=(dim, )),
a3=L.Bias(shape=(dim, )),
a4=L.Bias(shape=(dim, )),
b0=L.Scale(W_shape=(dim, )),
b1=L.Scale(W_shape=(dim, )),
b2=L.Scale(W_shape=(dim, )),
b3=L.Bias(shape=(dim, )),
)
def setUp(self):
self.x = numpy.random.uniform(-1, 1, (3, 2, 3)).astype(numpy.float32)
self.b = numpy.random.uniform(-1, 1, (2)).astype(numpy.float32)
self.y_expected = numpy.copy(self.x)
for i, j, k in numpy.ndindex(self.y_expected.shape):
self.y_expected[i, j, k] += self.b[j]
self.gy = numpy.random.uniform(-1, 1, (3, 2, 3)).astype(numpy.float32)
axis = 1
if self.learn_b:
self.link = links.Bias(axis, self.b.shape)
self.link.b.data = self.b
else:
self.link = links.Bias(axis, None)
self.link.cleargrads()
def __init__(self, blocks=5):
super(T2Resnet_multi, self).__init__()
self.blocks = blocks
with self.init_scope():
self.l1 = L.Convolution2D(in_channels=104,
out_channels=ch,
ksize=3,
pad=1)
for i in range(1, blocks):
block_wobn = Block(True)
block_bn = Block(False)
self.add_link('b_wobn{}'.format(i), block_wobn)
self.add_link('b_bn{}'.format(i), block_bn)
# policy network
self.policy = L.Convolution2D(
in_channels=ch,
out_channels=MOVE_DIRECTION_LABEL_NUM,
ksize=1,
nobias=True)
self.policy_bias = L.Bias(shape=(9 * 9 * MOVE_DIRECTION_LABEL_NUM))
# value network
self.value1 = L.Convolution2D(
in_channels=ch, out_channels=MOVE_DIRECTION_LABEL_NUM, ksize=1)
self.value1_bn = L.BatchNormalization(MOVE_DIRECTION_LABEL_NUM)
self.value2 = L.Linear(9 * 9 * MOVE_DIRECTION_LABEL_NUM, fcl)
self.value3 = L.Linear(fcl, 1)
def __init__(self, in_size=512, out_channel=3):
super(Generator, self).__init__()
with self.init_scope():
self.fc7_1 = L.Linear(in_size, 4 * 4 * 512, nobias=True)
self.b7_1 = L.Bias(shape=(512, ))
self.conv7_2 = L.Convolution2D(512, 512, ksize=3, pad=1)
self.conv6_1 = L.Convolution2D(512, 512, ksize=3, pad=1)
self.conv6_2 = L.Convolution2D(512, 512, ksize=3, pad=1)
self.conv5_1 = L.Convolution2D(512, 512, ksize=3, pad=1)
self.conv5_2 = L.Convolution2D(512, 512, ksize=3, pad=1)
self.conv4_1 = L.Convolution2D(512, 512, ksize=3, pad=1)
self.conv4_2 = L.Convolution2D(512, 512, ksize=3, pad=1)
self.conv3_1 = L.Convolution2D(512, 256, ksize=3, pad=1)
self.conv3_2 = L.Convolution2D(256, 256, ksize=3, pad=1)
self.conv2_1 = L.Convolution2D(256, 128, ksize=3, pad=1)
self.conv2_2 = L.Convolution2D(128, 128, ksize=3, pad=1)
self.conv1_1 = L.Convolution2D(128, 64, ksize=3, pad=1)
self.conv1_2 = L.Convolution2D(64, 64, ksize=3, pad=1)
self.conv0_0 = L.Convolution2D(64, out_channel, ksize=1)
def __init__(self,
in_dim,
hidden_dim,
node_dim,
n_hidden_layers=0,
activation=functions.relu,
dropout=0.0):
# in_dim should be layer_dim
# hidden_dim is edge_hidden_dim
# out_dim should be node_dim ** 2
super(EdgeNetwork, self).__init__()
with self.init_scope():
first_linear_layer = links.Linear(in_size=in_dim,
out_size=hidden_dim)
rest_linear_layers = [
links.Linear(in_size=hidden_dim, out_size=hidden_dim)
for _ in range(n_hidden_layers - 1)
]
hidden_layers = [first_linear_layer] + rest_linear_layers
self.hidden_layers = chainer.ChainList(*hidden_layers)
if n_hidden_layers == 0:
self.output_layer = links.Linear(in_size=in_dim,
out_size=node_dim**2)
else:
self.output_layer = links.Linear(in_size=hidden_dim,
out_size=node_dim**2)
self.bias_add_layer = links.Bias(axis=1)
self.in_dim = in_dim
self.hidden_dim = hidden_dim
self.node_dim = node_dim
self.out_dim = self.node_dim**2
self.n_hidden_layers = n_hidden_layers
self.activation = activation
self.dropout = dropout
def __init__(self,
n_features=None,
n_dim=8,
lossfun=F.mean_squared_error,
lambda0=5e-3,
lambda1=5e-3,
lambda2=5e-3,
init_bias=0.0,
intx_term=True,
total_nobs=1):
self.n_dim = n_dim
self.n_features = n_features
self.lossfun = lossfun
self.lambda0 = lambda0
self.lambda1 = lambda1
self.lambda2 = lambda2
self.intx_term = intx_term
self.total_nobs = total_nobs
# These are all the learned weights corresponding
# to the overall bias, slope per feature, and latent
# interaction vector per feature
super(FM, self).__init__(bias=L.Bias(shape=(1, )),
slope=L.EmbedID(n_features, 1),
latent=L.EmbedID(n_features, n_dim))
# Xavier initialize weights
c = np.sqrt(n_features * n_dim)
self.latent.W.data[...] = np.random.randn(n_features, n_dim) / c
d = np.sqrt(n_features)
self.slope.W.data[...] = np.random.randn(n_features, 1) / d
self.bias.b.data[...] *= 0.0
self.bias.b.data[...] += init_bias
def __init__(self, ch=512, ch_in=512, upsample=True, enable_blur=False):
super().__init__()
self.upsample = upsample
self.ch = ch
self.ch_in = ch_in
with self.init_scope():
self.b0 = L.Bias(axis=1, shape=(ch, ))
self.b1 = L.Bias(axis=1, shape=(ch, ))
self.n0 = NoiseBlock(ch)
self.n1 = NoiseBlock(ch)
self.c0 = EqualizedConv2d(ch_in, ch, 3, 1, 1, nobias=True)
self.c1 = EqualizedConv2d(ch, ch, 3, 1, 1, nobias=True)
self.blur_k = None
self.enable_blur = enable_blur
def __init__(self,
dim_in,
dim_hidden,
dim_latent,
num_layers,
num_trans,
temperature,
num_zsamples=1):
super(VAE, self).__init__()
# initialise first encoder and decoder hidden layer separately because
# the input and output dims differ from the other hidden layers
self.qlin0 = L.Linear(dim_in, dim_hidden)
self.plin0 = L.Linear(dim_latent, dim_hidden)
self._children.append('qlin0')
self._children.append('plin0')
for i in range(num_layers - 1):
# encoder
layer_name = 'qlin' + str(i + 1)
setattr(self, layer_name, L.Linear(2 * dim_hidden, dim_hidden))
self._children.append(layer_name)
# decoder
layer_name = 'plin' + str(i + 1)
setattr(self, layer_name, L.Linear(2 * dim_hidden, dim_hidden))
self._children.append(layer_name)
# initialise the encoder and decoder output layer separately because
# the input and output dims differ from the other hidden layers
self.qlin_mu = L.Linear(2 * dim_hidden, dim_latent)
self.qlin_ln_var = L.Linear(2 * dim_hidden, dim_latent)
self.plin_ln_var = L.Linear(2 * dim_hidden, dim_in)
self.plin_mu = L.Linear(2 * dim_hidden, dim_in)
self._children.append('qlin_mu')
self._children.append('qlin_ln_var')
self._children.append('plin_mu')
self._children.append('plin_ln_var')
# flow
for i in range(num_trans):
layer_name = 'flow_w_' + str(i) # weights
setattr(self, layer_name,
L.Scale(axis=1, W_shape=(dim_latent), bias_term=False))
self._children.append(layer_name)
layer_name = 'flow_b_' + str(i) # bias
setattr(self, layer_name, L.Bias(axis=0, shape=(1)))
self._children.append(layer_name)
layer_name = 'flow_u_' + str(i) # scaling factor u
setattr(self, layer_name,
L.Scale(axis=1, W_shape=(dim_latent), bias_term=False))
self._children.append(layer_name)
self.num_layers = num_layers
self.num_trans = num_trans
self.temperature = temperature
self.num_zsamples = num_zsamples
self.epochs_seen = 0
def __init__(self, blocks=20):
super(PolicyValueResnet, self).__init__()
self.blocks = blocks
with self.init_scope():
self.l1 = L.Convolution2D(in_channels=104,
out_channels=ch,
ksize=3,
pad=1)
self.l2 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
for i in range(1, 20):
self.add_link('b{}'.format(i), Block())
# policy network
self.policy = L.Convolution2D(
in_channels=ch,
out_channels=MOVE_DIRECTION_LABEL_NUM,
ksize=1,
nobias=True)
self.policy_bias = L.Bias(shape=(9 * 9 * MOVE_DIRECTION_LABEL_NUM))
# value network
self.value1 = L.Convolution2D(
in_channels=ch, out_channels=MOVE_DIRECTION_LABEL_NUM, ksize=1)
self.value2 = L.Linear(9 * 9 * MOVE_DIRECTION_LABEL_NUM, fcl)
self.value3 = L.Linear(fcl, 1)
def __init__(self, blocks=5):
super(PolicyValueResnet, self).__init__()
self.blocks = blocks
with self.init_scope():
self.l1 = L.Convolution2D(in_channels=104,
out_channels=ch,
ksize=3,
pad=1)
links = [("root0", RoopBlock(ch, ch))]
n_in = ch
n_out = 256
for index in range(1, 5, 1):
links += [("root{}".format(index), RoopBlock(n_in, n_out))]
n_in *= 1
n_out *= 1
for link in links:
self.add_link(*link)
self.forward = links
# policy network
self.policy = L.Convolution2D(
in_channels=n_in,
out_channels=MOVE_DIRECTION_LABEL_NUM,
ksize=1,
nobias=True)
self.policy_bias = L.Bias(shape=(9 * 9 * MOVE_DIRECTION_LABEL_NUM))
# value network
self.value1 = L.Convolution2D(
in_channels=n_in,
out_channels=MOVE_DIRECTION_LABEL_NUM,
ksize=1)
self.value2 = L.Linear(9 * 9 * MOVE_DIRECTION_LABEL_NUM, fcl)
self.value3 = L.Linear(fcl, 1)
def __init__(self, data, name, learnable=False, is_observed=False):
self._current_value = coerce_to_dtype(data, is_observed)
self.name = name
self._observed = is_observed
self.parents = ()
self.learnable = learnable
if learnable:
self.link = L.Bias(axis=1, shape=self._current_value.shape[1:])
def __init__(self):
super(PolicyNetwork, self).__init__()
with self.init_scope():
self.l1 = L.Convolution2D(in_channels=104,
out_channels=ch,
ksize=3,
pad=1)
self.l2 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l3 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l4 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l5 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l6 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l7 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l8 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l9 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l10 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l11 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l12 = L.Convolution2D(in_channels=ch,
out_channels=ch,
ksize=3,
pad=1)
self.l13 = L.Convolution2D(in_channels=ch,
out_channels=MOVE_DIRECTION_LABEL_NUM,
ksize=1,
nobias=True)
self.l13_bias = L.Bias(shape=(9 * 9 * MOVE_DIRECTION_LABEL_NUM))
def __init__(self, n_user: int, n_item: int, n_factor=10):
super(LinearWeightAdaption, self).__init__()
# settings
self.n_user = n_user
self.n_item = n_item
self.n_factor = n_factor
with self.init_scope():
# architecture
self.embedId = links.EmbedID(self.n_item, self.n_factor)
self.linear = links.Convolution2D(1, 1, (1, 2), nobias=True, pad=0)
self.bias = links.Bias(shape=(1, ))
def __init__(self):
ksize = 3
super(SLPolicy, self).__init__()
with self.init_scope():
self.block1 = Block(64, ksize)
self.block2 = Block(128, ksize)
self.block3 = Block(128, ksize)
self.block4 = Block(128, ksize)
self.block5 = Block(128, ksize)
self.block6 = Block(128, ksize)
self.block7 = Block(128, ksize)
self.block8 = Block(128, ksize)
self.conv9 = L.Convolution2D(128, 1, 1, nobias=True)
self.bias10 = L.Bias(shape=(64))
def deconv(self, variable):
v = variable
if (v.creator is not None):
# Convolution -> Deconvolutionに変換
if (v.creator.label == 'Convolution2DFunction'):
print(v.creator.label, v.rank)
convW = v.creator.inputs[1].data
in_cn, out_cn = convW.shape[0], convW.shape[1] # in/out channels
kh, kw = convW.shape[2], convW.shape[3] # kernel size
sx, sy = v.creator.sx, v.creator.sy # stride
pw, ph = v.creator.pw, v.creator.ph # padding
name = 'conv' + v.rank # temporal layer name
super(DeconvNet, self).add_link(
name,
L.Deconvolution2D(in_cn,
out_cn, (kh, kw),
stride=(sy, sx),
pad=(ph, pw),
nobias=True,
initialW=convW))
self.forwards[name] = self[name]
# もし畳み込み層にバイアスがある場合、それも登録
if len(v.creator.inputs) == 3:
F.bias(v)
b = v.creator.inputs[2].data
bname = 'convb' + v.rank
super(DeconvNet, self).add_link(bname, L.Bias(shape=b.shape))
self[bname].b.data = b
self.depends[bname] = (parent)
self.depends[name] = (bname)
self.forwards[bname] = self[bname]
self.layers.append((bname, [parent], name))
else:
self.depends[name] = (parent)
elif (v.creator.label == 'ReLU'):
name = parent
elif (v.creator.label == 'MaxPooling2D'):
kw, kh = v.creator.kw, v.creator.kh
sx, sy = v.creator.sx, v.creator.sy
pw, ph = v.creator.pw, v.creator.ph
name = 'maxpool' + v.rank
self.depends[name] = (parent)
self.forwards[name] = lambda x: F.unpooling_2d(
x, (kh, kw), stride=(sy, sx), pad=(ph, pw))
self.register_inv_layer(v.creator.inputs[0], name)
else:
depends['output'] = parent
def darknetConv2D(in_channel, out_channel, bn=True):
if (bn):
return Chain(
c=L.Convolution2D(in_channel,
out_channel,
ksize=3,
pad=1,
nobias=True),
n=L.BatchNormalization(out_channel, use_beta=False, eps=0.000001),
b=L.Bias(shape=[
out_channel,
]),
)
else:
return Chain(
c=L.Convolution2D(in_channel,
out_channel,
ksize=3,
pad=1,
nobias=True),
b=L.Bias(shape=[
out_channel,
]),
)
def __init__(self,
hidden_dim,
tracking_lstm_hidden_dim,
num_actions=2,
make_logits=False):
super(TrackingLSTM, self).__init__(
W_x=L.Linear(tracking_lstm_hidden_dim / 2,
tracking_lstm_hidden_dim / 2 * 4),
W_h=L.Linear(tracking_lstm_hidden_dim / 2,
tracking_lstm_hidden_dim / 2 * 4),
bias=L.Bias(axis=1, shape=(tracking_lstm_hidden_dim / 2 * 4, )),
)
# TODO: TrackingLSTM should be able to be a size different from hidden_dim.
if make_logits:
self.add_link('logits',
L.Linear(tracking_lstm_hidden_dim / 2, num_actions))
def __init__(self, n_classes_fcn, n_classes_yolo, n_boxes):
super(YOLOv2, self).__init__(
conv1=L.Convolution2D(3, 64, 3, stride=1, pad=1, nobias=True),
bn1=L.BatchNormalization(64, use_beta=False, eps=2e-5),
bias1=L.Bias(shape=(64, )),
conv2=L.Convolution2D(None, 64, 3, stride=1, pad=1, nobias=True),
bn2=L.BatchNormalization(64, use_beta=False, eps=2e-5),
bias2=L.Bias(shape=(64, )),
conv3=L.Convolution2D(None, 128, 3, stride=1, pad=1, nobias=True),
bn3=L.BatchNormalization(128, use_beta=False, eps=2e-5),
bias3=L.Bias(shape=(128, )),
conv4=L.Convolution2D(None, 128, 3, stride=1, pad=1, nobias=True),
bn4=L.BatchNormalization(128, use_beta=False, eps=2e-5),
bias4=L.Bias(shape=(128, )),
conv5=L.Convolution2D(None, 256, 3, stride=1, pad=1, nobias=True),
bn5=L.BatchNormalization(256, use_beta=False, eps=2e-5),
bias5=L.Bias(shape=(256, )),
conv6=L.Convolution2D(None, 256, 3, stride=1, pad=1, nobias=True),
bn6=L.BatchNormalization(256, use_beta=False, eps=2e-5),
bias6=L.Bias(shape=(256, )),
conv7=L.Convolution2D(None, 256, 3, stride=1, pad=1, nobias=True),
bn7=L.BatchNormalization(256, use_beta=False, eps=2e-5),
bias7=L.Bias(shape=(256, )),
pool1=L.Convolution2D(None, n_classes_fcn, 1, stride=1, pad=0),
upsample3=L.Deconvolution2D(None,
n_classes_fcn,
ksize=16,
stride=8,
pad=4),
conv14=L.Convolution2D(None, 1024, 3, stride=1, pad=1,
nobias=True),
bn14=L.BatchNormalization(1024, use_beta=False, eps=2e-5),
bias14=L.Bias(shape=(1024, )),
conv15=L.Convolution2D(None,
n_boxes * (5 + n_classes_yolo),
ksize=1,
stride=1,
pad=0),
)
self.n_boxes = n_boxes
self.n_classes_fcn = n_classes_fcn
self.n_classes_yolo = n_classes_yolo
self.finetune = False
def __init__(self, n_features=None, n_dim=8, lossfun=F.mean_squared_error,
lambda0=1, lambda1=1, lambda2=1, init_bias_mu=0.0,
init_bias_lv=0.0, intx_term=True, total_nobs=1):
self.n_dim = n_dim
self.n_features = n_features
self.lossfun = lossfun
self.lambda0 = lambda0
self.lambda1 = lambda1
self.lambda2 = lambda2
self.intx_term = intx_term
self.total_nobs = total_nobs
# In contrast to the FM model, the slopes and latent vectors
# will have means (mu) and log variances (lv) for each component.
ones_3d = (1, 1, 1)
super(AutoVFM, self).__init__(bias_mu=L.Bias(shape=(1,)),
bias_lv=L.Bias(shape=(1,)),
slop_mu=L.Bias(shape=(1, 1)),
slop_lv=L.Bias(shape=(1, 1)),
slop_delta_mu=L.EmbedID(n_features, 1,
ignore_label=-1),
slop_delta_lv=L.EmbedID(n_features, 1,
ignore_label=-1),
feat_mu_vec=L.Bias(shape=(1, 1, n_dim)),
feat_lv_vec=L.Bias(shape=(1, 1, n_dim)),
hyper_feat_lv_vec=L.Bias(shape=ones_3d),
feat_delta_mu=L.EmbedID(n_features, n_dim,
ignore_label=-1),
feat_delta_lv=L.EmbedID(n_features, n_dim,
ignore_label=-1),
hyper_feat_delta_lv=L.Bias(shape=ones_3d))
# Xavier initialize weights
c = np.sqrt(n_features * n_dim) * 1e3
d = np.sqrt(n_features) * 1e3
self.feat_delta_mu.W.data[...] = np.random.randn(n_features, n_dim) / c
self.feat_delta_lv.W.data[...] = np.random.randn(n_features, n_dim) / c
self.slop_delta_mu.W.data[...] = np.random.randn(n_features, 1) / d
self.slop_delta_lv.W.data[...] = np.random.randn(n_features, 1) / d
self.bias_mu.b.data[...] *= 0.0
self.bias_mu.b.data[...] += init_bias_mu
self.bias_lv.b.data[...] *= 0.0
self.bias_lv.b.data[...] += init_bias_lv
def get_conv_stack(self,
input,
output,
ksize=3,
stride=1,
pad=1,
nobias=True,
use_beta=False):
conv = L.Convolution2D(input,
output,
ksize=ksize,
stride=stride,
pad=pad,
nobias=nobias)
bn = L.BatchNormalization(output, use_beta=use_beta)
bias = L.Bias(shape=(output, ))
return conv, bn, bias
def __init__(self):
super(PolicyNetwork, self).__init__()
with self.init_scope():
self.conv1_1 = L.Convolution2D(104, ch, 3, pad=1)
self.conv1_2 = L.Convolution2D(ch, ch, 3, pad=1)
self.conv2_1 = L.Convolution2D(ch, ch * 2, 3, pad=1)
self.conv2_2 = L.Convolution2D(ch * 2, ch * 2, 3, pad=1)
self.conv3_1 = L.Convolution2D(ch * 2, ch * 4, 3, pad=1)
self.conv3_2 = L.Convolution2D(ch * 4, ch * 4, 3, pad=1)
self.conv3_3 = L.Convolution2D(ch * 4, ch * 4, 3, pad=1)
self.conv3_4 = L.Convolution2D(ch * 4, ch * 4, 3, pad=1)
self.l13 = L.Convolution2D(in_channels=ch * 4,
out_channels=MOVE_DIRECTION_LABEL_NUM,
ksize=1,
nobias=True)
self.l13_bias = L.Bias(shape=(9 * 9 * MOVE_DIRECTION_LABEL_NUM))
def __init__(self,
hidden_dim,
tracking_lstm_hidden_dim,
use_external=False,
prefix="TreeLSTMChain",
gpu=-1):
super(TreeLSTMChain, self).__init__(
W_l=L.Linear(hidden_dim / 2, hidden_dim / 2 * 5, nobias=True),
W_r=L.Linear(hidden_dim / 2, hidden_dim / 2 * 5, nobias=True),
b=L.Bias(axis=1, shape=(hidden_dim / 2 * 5, )),
)
assert hidden_dim % 2 == 0, "The hidden_dim must be even because contains c and h."
self.hidden_dim = hidden_dim / 2
self.__gpu = gpu
self.__mod = cuda.cupy if gpu >= 0 else np
self.use_external = use_external
if use_external:
self.add_link(
'W_external',
L.Linear(tracking_lstm_hidden_dim / 2, hidden_dim / 2 * 5))
def __init__(self):
super(MyChain, self).__init__(
l1=L.Convolution2D(in_channels=None,
out_channels=k,
ksize=3,
pad=1),
l2=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l3=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l4=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l5=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l6=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l7=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l8=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l9=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l10=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l11=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l12=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l13=L.Convolution2D(in_channels=k,
out_channels=len(shogi.PIECE_TYPES),
ksize=1,
nobias=True),
l13_2=L.Bias(shape=(9 * 9 * len(shogi.PIECE_TYPES))))
def __init__(self):
super(PolicyNetwork, self).__init__(
l1=L.Convolution2D(in_channels=None,
out_channels=k,
ksize=3,
pad=1),
l2=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l3=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l4=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l5=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l6=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l7=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l8=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l9=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l10=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l11=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l12=L.Convolution2D(in_channels=k, out_channels=k, ksize=3, pad=1),
l13=L.Convolution2D(in_channels=k,
out_channels=MOVE_DIRECTION_LABEL_NUM,
ksize=1,
nobias=True),
l13_2=L.Bias(shape=(9 * 9 * MOVE_DIRECTION_LABEL_NUM)))
def __init__(self):
self.n_experts = 128
self.z_dim = 10
self.data_dim = 28 * 28
super(DDGMNet, self).__init__(
# Generative Model
gm_linear1=L.Linear(self.z_dim, 128),
gm_linear2=L.Linear(128, 128),
gm_linear3=L.Linear(128, self.data_dim),
# Energy Model
# add em_ prefix to components of the energy model for a trick about gradients
em_linear1=L.Linear(self.data_dim, 128),
em_linear2=L.Linear(128, 128),
em_experts=L.Linear(128, self.n_experts),
em_bias=L.Bias(shape=(self.data_dim, )),
)
self.add_param('em_ln_var', tuple())
self.em_ln_var.data = self.xp.zeros(self.em_ln_var.data.shape,
dtype='float32')
self.em_params = []
for k, v in self.namedparams():
if k.startswith('/em_'): self.em_params.append(v)
def __init__(self, in_channels, out_channels,
ksize=None, stride=1, pad=0,
batch_normalize=False, activation='linear', initialW=None):
super(Convolution, self).__init__()
self.batch_normalize = batch_normalize
if activation not in ['linear', 'leaky']:
raise ValueError()
if 'linear' == activation:
self.activation = F.identity
elif 'leaky' == activation:
self.activation = lambda x: F.leaky_relu(x, slope=0.1)
with self.init_scope():
self.conv = L.Convolution2D(in_channels, out_channels,
ksize, stride, pad,
nobias=True,
initialW=initialW)
if batch_normalize:
self.bn = L.BatchNormalization(out_channels)
self.b = L.Bias(shape=(out_channels,))
def __init__(self, n_classes, n_boxes):
super(YOLOv2, self).__init__(
##### common layers for both pretrained layers and yolov2 #####
conv1=L.Convolution2D(3, 32, ksize=3, stride=1, pad=1,
nobias=True),
bn1=L.BatchNormalization(32, use_beta=False, eps=2e-5),
bias1=L.Bias(shape=(32, )),
conv2=L.Convolution2D(32,
64,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn2=L.BatchNormalization(64, use_beta=False, eps=2e-5),
bias2=L.Bias(shape=(64, )),
conv3=L.Convolution2D(64,
128,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn3=L.BatchNormalization(128, use_beta=False, eps=2e-5),
bias3=L.Bias(shape=(128, )),
conv4=L.Convolution2D(128,
64,
ksize=1,
stride=1,
pad=0,
nobias=True),
bn4=L.BatchNormalization(64, use_beta=False, eps=2e-5),
bias4=L.Bias(shape=(64, )),
conv5=L.Convolution2D(64,
128,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn5=L.BatchNormalization(128, use_beta=False, eps=2e-5),
bias5=L.Bias(shape=(128, )),
conv6=L.Convolution2D(128,
256,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn6=L.BatchNormalization(256, use_beta=False, eps=2e-5),
bias6=L.Bias(shape=(256, )),
conv7=L.Convolution2D(256,
128,
ksize=1,
stride=1,
pad=0,
nobias=True),
bn7=L.BatchNormalization(128, use_beta=False, eps=2e-5),
bias7=L.Bias(shape=(128, )),
conv8=L.Convolution2D(128,
256,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn8=L.BatchNormalization(256, use_beta=False, eps=2e-5),
bias8=L.Bias(shape=(256, )),
conv9=L.Convolution2D(256,
512,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn9=L.BatchNormalization(512, use_beta=False, eps=2e-5),
bias9=L.Bias(shape=(512, )),
conv10=L.Convolution2D(512,
256,
ksize=1,
stride=1,
pad=0,
nobias=True),
bn10=L.BatchNormalization(256, use_beta=False, eps=2e-5),
bias10=L.Bias(shape=(256, )),
conv11=L.Convolution2D(256,
512,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn11=L.BatchNormalization(512, use_beta=False, eps=2e-5),
bias11=L.Bias(shape=(512, )),
conv12=L.Convolution2D(512,
256,
ksize=1,
stride=1,
pad=0,
nobias=True),
bn12=L.BatchNormalization(256, use_beta=False, eps=2e-5),
bias12=L.Bias(shape=(256, )),
conv13=L.Convolution2D(256,
512,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn13=L.BatchNormalization(512, use_beta=False, eps=2e-5),
bias13=L.Bias(shape=(512, )),
conv14=L.Convolution2D(512,
1024,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn14=L.BatchNormalization(1024, use_beta=False, eps=2e-5),
bias14=L.Bias(shape=(1024, )),
conv15=L.Convolution2D(1024,
512,
ksize=1,
stride=1,
pad=0,
nobias=True),
bn15=L.BatchNormalization(512, use_beta=False, eps=2e-5),
bias15=L.Bias(shape=(512, )),
conv16=L.Convolution2D(512,
1024,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn16=L.BatchNormalization(1024, use_beta=False, eps=2e-5),
bias16=L.Bias(shape=(1024, )),
conv17=L.Convolution2D(1024,
512,
ksize=1,
stride=1,
pad=0,
nobias=True),
bn17=L.BatchNormalization(512, use_beta=False, eps=2e-5),
bias17=L.Bias(shape=(512, )),
conv18=L.Convolution2D(512,
1024,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn18=L.BatchNormalization(1024, use_beta=False, eps=2e-5),
bias18=L.Bias(shape=(1024, )),
###### new layer
conv19=L.Convolution2D(1024,
1024,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn19=L.BatchNormalization(1024, use_beta=False),
bias19=L.Bias(shape=(1024, )),
conv20=L.Convolution2D(1024,
1024,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn20=L.BatchNormalization(1024, use_beta=False),
bias20=L.Bias(shape=(1024, )),
conv21=L.Convolution2D(3072,
1024,
ksize=3,
stride=1,
pad=1,
nobias=True),
bn21=L.BatchNormalization(1024, use_beta=False),
bias21=L.Bias(shape=(1024, )),
conv22=L.Convolution2D(1024,
n_boxes * (5 + n_classes),
ksize=1,
stride=1,
pad=0,
nobias=True),
bias22=L.Bias(shape=(n_boxes * (5 + n_classes), )),
)
self.train = False
self.finetune = False
self.n_boxes = n_boxes
self.n_classes = n_classes