def create_initializer(init_type, scale=None, fillvalue=None):
if init_type == 'identity':
return initializers.Identity() if scale is None else initializers.Identity(scale=scale)
if init_type == 'constant':
return initializers.Constant(fillvalue)
if init_type == 'zero':
return initializers.Zero()
if init_type == 'one':
return initializers.One()
if init_type == 'normal':
return initializers.Normal() if scale is None else initializers.Normal(scale)
if init_type == 'glorotNormal':
return initializers.GlorotNormal() if scale is None else initializers.GlorotNormal(scale)
if init_type == 'heNormal':
return initializers.HeNormal() if scale is None else initializers.HeNormal(scale)
if init_type == 'orthogonal':
return initializers.Orthogonal(
scale) if scale is None else initializers.Orthogonal(scale)
if init_type == 'uniform':
return initializers.Uniform(
scale) if scale is None else initializers.Uniform(scale)
if init_type == 'leCunUniform':
return initializers.LeCunUniform(
scale) if scale is None else initializers.LeCunUniform(scale)
if init_type == 'glorotUniform':
return initializers.GlorotUniform(
scale) if scale is None else initializers.GlorotUniform(scale)
if init_type == 'heUniform':
return initializers.HeUniform(
scale) if scale is None else initializers.HeUniform(scale)
raise ValueError("Unknown initializer type: {0}".format(init_type))
def __init__(self, num_entities, num_relations, fast_eval, embedding_dim=200):
super(DistMult, self).__init__()
self.fast_eval = fast_eval
self.num_entities = num_entities
self.num_relations = num_relations
self.embedding_dim = embedding_dim
with self.init_scope():
self.emb_e = L.EmbedID(
num_entities, embedding_dim, initialW=I.GlorotNormal())
self.emb_rel = L.EmbedID(
num_relations, embedding_dim, initialW=I.GlorotNormal())
def __init__(self, ch, k, l, m, n):
super(Reduction_A, self).__init__()
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn2_1 = ConvBN(ch,
n,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
self.convbn3_1 = ConvBN(ch,
k,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_2 = ConvBN(k,
l,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
self.convbn3_3 = ConvBN(l,
m,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
def __init__(self, num_entities, num_relations, fast_eval):
super(ConvE, self).__init__()
self.fast_eval = fast_eval
self.num_entities = num_entities
self.num_relations = num_relations
self.embedding_dim = 200
with self.init_scope():
self.emb_e = L.EmbedID(
num_entities, self.embedding_dim, initialW=I.GlorotNormal())
self.emb_rel = L.EmbedID(
num_relations, self.embedding_dim, initialW=I.GlorotNormal())
self.conv1 = L.Convolution2D(1, 32, 3, stride=1, pad=0)
self.bias = L.EmbedID(num_entities, 1)
self.fc = L.Linear(10368, self.embedding_dim)
self.bn0 = L.BatchNormalization(1)
self.bn1 = L.BatchNormalization(32)
self.bn2 = L.BatchNormalization(self.embedding_dim)
def __call__(self, atoms_1, g_1, atoms_2, g_2):
"""
:param atoms_1: atomic representation of molecule 1, with shape of (mb, N_1, hidden_dim)
:param g_1: molecular representation of molecule 1, with shape of (mb, out_dim)
:param atoms_2: atomic representation of molecule 2, with shape of (mb, N_2, hidden_dim)
:param g_2: molecular representation of molecule 2, with shape of (mb, out_dim)
:return:
"""
mb = atoms_1.shape[0]
N_1 = atoms_1.shape[1]
N_2 = atoms_2.shape[1]
# C: (mb, N_2, N_1)
C = self.compute_attention(query=atoms_2, key=atoms_1)
W1_data = self.xp.zeros(shape=(N_2, self.out_dim))
if self.xp == cuda.cupy:
W1_data = cuda.to_gpu(W1_data)
initializers.GlorotNormal()(W1_data)
self.W1 = chainer.Parameter(initializer=W1_data,
shape=(N_2, self.out_dim),
name="W1")
self.register_persistent("W1")
W2_data = self.xp.zeros(shape=(N_1, self.out_dim))
if self.xp == cuda.cupy:
W2_data = cuda.to_gpu(W2_data)
initializers.GlorotNormal()(W2_data)
self.W2 = chainer.Parameter(initializer=W2_data,
shape=(N_1, self.out_dim),
name="W2")
self.register_persistent("W2")
# attn_1: (mb, N_1, out_dim)
attn_1 = functions.matmul(
functions.transpose(C, (0, 2, 1)),
functions.tile(functions.expand_dims(self.W1, axis=0),
reps=(mb, 1, 1)))
# attn_2: (mb, N_2, out_dim)
attn_2 = functions.matmul(
C,
functions.tile(functions.expand_dims(self.W2, axis=0),
reps=(mb, 1, 1)))
# compact_1: (mb, out_dim)
compact_1 = functions.sum(attn_1 * self.j_layer(atoms_1), axis=1)
# compact_2: (mb, out_dim)
compact_2 = functions.sum(attn_2 * self.j_layer(atoms_2), axis=1)
return compact_1, compact_2
def __init__(self, base, n_base_output, scales):
super().__init__()
with self.init_scope():
self.base = base
self.inner = chainer.ChainList()
self.outer = chainer.ChainList()
init = {'initialW': initializers.GlorotNormal()}
for _ in range(n_base_output):
self.inner.append(L.Convolution2D(256, 1, **init))
self.outer.append(L.Convolution2D(256, 3, pad=1, **init))
self.scales = scales
def __init__(self, ch, scale=1.0):
super(Inception_Resnet_A, self).__init__()
self.scale = scale
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn1 = ConvBN(ch,
32,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn2_1 = ConvBN(ch,
32,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn2_2 = ConvBN(32,
32,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
self.convbn3_1 = ConvBN(ch,
32,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_2 = ConvBN(32,
48,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
self.convbn3_3 = ConvBN(48,
64,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
self.conv4 = L.Convolution2D(128,
ch,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
def __init__(self, ch):
super(Reduction_B, self).__init__()
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn2_1 = ConvBN(ch,
256,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn2_2 = ConvBN(256,
384,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
self.convbn3_1 = ConvBN(ch,
256,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_2 = ConvBN(256,
288,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
self.convbn4_1 = ConvBN(ch,
256,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn4_2 = ConvBN(256,
288,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
self.convbn4_3 = ConvBN(288,
320,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
def __init__(self, ch):
super(Inception_A, self).__init__()
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn1 = ConvBN(ch,
96,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn2_1 = ConvBN(ch,
96,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_1 = ConvBN(ch,
64,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_2 = ConvBN(64,
96,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
self.convbn4_1 = ConvBN(ch,
64,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn4_2 = ConvBN(64,
96,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
self.convbn4_3 = ConvBN(96,
96,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
def __init__(self, ch):
super(Stem, self).__init__()
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn1 = ConvBN(ch, 32, ksize=3, stride=2,
initialW=initialW, pad=0)
self.convbn2 = ConvBN(32, 32, ksize=3, stride=1,
initialW=initialW, pad=0)
self.convbn3 = ConvBN(32, 64, ksize=3, stride=1,
initialW=initialW, pad=1)
self.convbn4 = ConvBN(64, 80, ksize=1, stride=1,
initialW=initialW, pad=0)
self.convbn5 = ConvBN(80, 192, ksize=3, stride=1,
initialW=initialW, pad=0)
self.convbn6 = ConvBN(192, 256, ksize=3, stride=2,
initialW=initialW, pad=0)
def __init__(self, base, n_base_output, scales):
super(FPN, self).__init__()
with self.init_scope():
self.base = base
self.inner = chainer.ChainList()
self.outer = chainer.ChainList()
init = {'initialW': initializers.GlorotNormal()}
for _ in range(n_base_output):
self.inner.append(L.Convolution2D(256, 1, **init))
self.outer.append(L.Convolution2D(256, 3, pad=1, **init))
self.scales = scales
# hacks
self.n_base_output = n_base_output
self.n_base_output_minus1 = n_base_output - 1
self.scales_minus_n_base_output = len(scales) - n_base_output
def __init__(self, ch, scale):
super(Inception_Resnet_C, self).__init__()
self.scale = scale
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn1 = ConvBN(ch, 192, ksize=1, stride=1,
initialW=initialW, pad=0)
self.convbn2_1 = ConvBN(ch, 192, ksize=1, stride=1,
initialW=initialW, pad=0)
self.convbn2_2 = ConvBN(192, 192, ksize=(1, 3), stride=1,
initialW=initialW, pad=(0, 1))
self.convbn2_3 = ConvBN(192, 192, ksize=(3, 1), stride=1,
initialW=initialW, pad=(1, 0))
self.conv3 = L.Convolution2D(384, ch, ksize=1, stride=1,
initialW=initialW, pad=0)
def __init__(self,
in_size,
n_units,
dropout,
activation=F.tanh,
activation_on_final=False):
n_units = [
in_size,
] + n_units
params = {}
for i, n in enumerate(pairwise(n_units)):
# Initial W according the paper
params['l' + str(i)] = L.Linear(*n, initialW=I.GlorotNormal())
super(MLP, self).__init__(**params)
self.add_persistent('_dropout', dropout)
self.add_persistent('_activation_on_final', activation_on_final)
self._activation = activation
def __init__(self, ch, scale=1.):
super(Inception_Resnet_B, self).__init__()
self.scale = scale
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn1 = ConvBN(ch, 128, ksize=1, stride=1,
initialW=initialW, pad=0)
self.convbn2_1 = ConvBN(ch, 128, ksize=1, stride=1,
initialW=initialW, pad=0)
self.convbn2_2 = ConvBN(128, 128, ksize=(1, 7), stride=1,
initialW=initialW, pad=(0, 3))
self.convbn2_3 = ConvBN(128, 128, ksize=(7, 1), stride=1,
initialW=initialW, pad=(3, 0))
self.conv3 = L.Convolution2D(256, ch, ksize=1, stride=1,
initialW=initialW, pad=0)
def __init__(self, ch):
super(Reduction_B, self).__init__()
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn2_1 = ConvBN(ch,
192,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn2_2 = ConvBN(192,
192,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
self.convbn3_1 = ConvBN(ch,
256,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_2 = ConvBN(256,
256,
ksize=(1, 7),
stride=1,
initialW=initialW,
pad=(0, 3))
self.convbn3_3 = ConvBN(256,
320,
ksize=(7, 1),
stride=1,
initialW=initialW,
pad=(3, 0))
self.convbn3_4 = ConvBN(320,
320,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
def __init__(self, ch):
super(Stem, self).__init__()
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn1 = ConvBN(ch,
32,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
self.convbn2 = ConvBN(32,
32,
ksize=3,
stride=1,
initialW=initialW,
pad=0)
self.convbn3 = ConvBN(32,
64,
ksize=3,
stride=1,
initialW=initialW,
pad=1)
self.convbn4 = ConvBN(64,
96,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
self.convbn5_1x1_1 = ConvBN(160,
64,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn5_7x1 = ConvBN(64,
64,
ksize=(7, 1),
stride=1,
initialW=initialW,
pad=(3, 0))
self.convbn5_1x7 = ConvBN(64,
64,
ksize=(1, 7),
stride=1,
initialW=initialW,
pad=(0, 3))
self.convbn5_3x3_1 = ConvBN(64,
96,
ksize=3,
stride=1,
initialW=initialW,
pad=0)
self.convbn5_1x1_2 = ConvBN(160,
64,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn5_3x3_2 = ConvBN(64,
96,
ksize=3,
stride=1,
initialW=initialW,
pad=0)
self.convbn6 = ConvBN(192,
192,
ksize=3,
stride=2,
initialW=initialW,
pad=0)
def initialize_parameters(self):
G_init = initializers.GlorotNormal()
#initializers.init_weight(self.W_predict.W.data, G_init)
initializers.init_weight(self.W_candidate.W.data, G_init)
self.initialize_LSTM(self.LSTM, G_init)
def setUp(self):
self.initializer = initializers.GlorotNormal(scale=0.1)
def __init__(self, ch):
super(Inception_C, self).__init__()
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn1 = ConvBN(ch,
256,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn2_1 = ConvBN(ch,
256,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_1 = ConvBN(ch,
384,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_2_1 = ConvBN(384,
256,
ksize=(1, 3),
stride=1,
initialW=initialW,
pad=(0, 1))
self.convbn3_2_2 = ConvBN(384,
256,
ksize=(3, 1),
stride=1,
initialW=initialW,
pad=(1, 0))
self.convbn4_1 = ConvBN(ch,
384,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn4_2 = ConvBN(384,
448,
ksize=(1, 3),
stride=1,
initialW=initialW,
pad=(0, 1))
self.convbn4_3 = ConvBN(448,
512,
ksize=(3, 1),
stride=1,
initialW=initialW,
pad=(1, 0))
self.convbn4_4_1 = ConvBN(512,
256,
ksize=(3, 1),
stride=1,
initialW=initialW,
pad=(1, 0))
self.convbn4_4_2 = ConvBN(512,
256,
ksize=(1, 3),
stride=1,
initialW=initialW,
pad=(0, 1))
def __init__(self, ch):
super(Inception_B, self).__init__()
initialW = initializers.GlorotNormal()
with self.init_scope():
self.convbn1 = ConvBN(ch,
128,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn2_1 = ConvBN(ch,
384,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_1 = ConvBN(ch,
192,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn3_2 = ConvBN(192,
224,
ksize=(1, 7),
stride=1,
initialW=initialW,
pad=(0, 3))
self.convbn3_3 = ConvBN(224,
256,
ksize=(7, 1),
stride=1,
initialW=initialW,
pad=(3, 0))
self.convbn4_1 = ConvBN(ch,
192,
ksize=1,
stride=1,
initialW=initialW,
pad=0)
self.convbn4_2 = ConvBN(192,
192,
ksize=(1, 7),
stride=1,
initialW=initialW,
pad=(0, 3))
self.convbn4_3 = ConvBN(192,
224,
ksize=(7, 1),
stride=1,
initialW=initialW,
pad=(3, 0))
self.convbn4_4 = ConvBN(224,
224,
ksize=(1, 7),
stride=1,
initialW=initialW,
pad=(0, 3))
self.convbn4_5 = ConvBN(224,
256,
ksize=(7, 1),
stride=1,
initialW=initialW,
pad=(3, 0))
