def __init__(self, dim_in, dim_out, dropout=0., act='relu', \
order=1, aggr='mean', is_train=True, bias='norm', **kwargs):
super(HighOrderAggregator, self).__init__(**kwargs)
self.dropout = dropout
self.bias = bias
self.act = F_ACT[act]
self.order = order
self.aggr = aggr
self.is_train = is_train
if dim_out > 0:
with tf.variable_scope(self.name + '_vars'):
for o in range(self.order + 1):
_k = 'order{}_weights'.format(o)
self.vars[_k] = glorot([dim_in, dim_out], name=_k)
for o in range(self.order + 1):
_k = 'order{}_bias'.format(o)
self.vars[_k] = zeros([dim_out], name=_k)
if self.bias == 'norm':
for o in range(self.order + 1):
_k1 = 'order{}_offset'.format(o)
_k2 = 'order{}_scale'.format(o)
self.vars[_k1] = zeros([1, dim_out], name=_k1)
self.vars[_k2] = ones([1, dim_out], name=_k2)
print('>> layer {}, dim: [{},{}]'.format(self.name, dim_in, dim_out))
if self.logging:
self._log_vars()
self.dim_in = dim_in
self.dim_out = dim_out
def __init__(self,
dim_in,
dim_out,
dropout=0.,
act='relu',
order=1,
aggr='mean',
is_train=True,
bias='norm',
**kwargs):
assert order <= 1, "now only support attention for order 0/1 layers"
super(AttentionAggregator, self).__init__(**kwargs)
self.dropout = dropout
self.bias = bias
self.act = F_ACT[act]
self.order = order
self.aggr = aggr
self.is_train = is_train
if 'mulhead' in kwargs.keys():
self.mulhead = int(kwargs['mulhead'])
else:
self.mulhead = 1
with tf.variable_scope(self.name + '_vars'):
self.vars['order0_weights'] = glorot([dim_in, dim_out],
name='order0_weights')
for k in range(self.mulhead):
self.vars['order1_weights_h{}'.format(k)] = glorot(
[dim_in, int(dim_out / self.mulhead)],
name='order1_weights_h{}'.format(k))
self.vars['order0_bias'] = zeros([dim_out], name='order0_bias')
for k in range(self.mulhead):
self.vars['order1_bias_h{}'.format(k)] = zeros(
[int(dim_out / self.mulhead)],
name='order1_bias_h{}'.format(k))
if self.bias == 'norm':
for o in range(self.order + 1):
_k1 = 'order{}_offset'.format(o)
_k2 = 'order{}_scale'.format(o)
self.vars[_k1] = zeros([1, dim_out], name=_k1)
self.vars[_k2] = ones([1, dim_out], name=_k2)
for k in range(self.mulhead):
self.vars['attention_0_h{}'.format(k)] = glorot(
[1, int(dim_out / self.mulhead)],
name='attention_0_h{}'.format(k))
self.vars['attention_1_h{}'.format(k)] = glorot(
[1, int(dim_out / self.mulhead)],
name='attention_1_h{}'.format(k))
self.vars['att_bias_0_h{}'.format(k)] = zeros(
[1], name='att_bias_0_h{}'.format(k))
self.vars['att_bias_1_h{}'.format(k)] = zeros(
[1], name='att_bias_1_h{}'.format(k))
print('>> layer {}, dim: [{},{}]'.format(self.name, dim_in, dim_out))
if self.logging:
self._log_vars()
self.dim_in = dim_in
self.dim_out = dim_out
def __init__(self, arch_gcn, dim_input_jk, mode=None, **kwargs):
"""
"""
super(JumpingKnowledge,self).__init__(**kwargs)
self.mode = mode
if not mode:
return
self.act = F_ACT[arch_gcn['act']]
self.bias = arch_gcn['bias']
self.dim_in = dim_input_jk
self.dim_out = arch_gcn['dim']
with tf.variable_scope(self.name + '_vars'):
self.vars['weights'] = glorot([self.dim_in,self.dim_out],name='weights')
self.vars['bias'] = zeros([self.dim_out],name='bias')
if self.bias == 'norm':
self.vars['offset'] = zeros([1,self.dim_out],name='offset')
self.vars['scale'] = ones([1,self.dim_out],name='scale')