def __init__(self, graph, r, name, bidirectional=False):
''' this defines a set of variables for the
augmentation portion of the graph. If graph
is not None, these variables get added to
graph. r is the dimension of the various
parameters. This includes variables for
both the backward and forward passes.
This creates the following variables:
no_parent, no_left_sibling, no_right_sibling
In theory I could add an extra set of the above
for each of the special symbols. I won't now,
though, although I may chnage that later.
'''
self.no_parent = nn.VariableNode([r], None, name=name + 'no_parent')
self.no_left_sibling = nn.VariableNode([r],
None,
name=name + 'no_left_sibling')
self.vs = [self.no_parent, self.no_left_sibling]
if bidirectional:
self.no_right_sibling = nn.VariableNode([r],
None,
name=name +
'no_right_sibling')
self.vs.append(self.no_right_sibling)
self.rvs = []
return
def __init__(self, config):
DefaultVariables.__init__(self, config)
# define the specific variables for this section
r = self.config.p.r
out_layers = self.config.p.out_layers
# the parameters for the input section
self.hyp_gru_block = self.add_GRUb_block('hyp', bidirectional = self.config.p.bidirectional)
self.statement_gru_block = self.add_GRUb_block('statement', bidirectional = self.config.p.bidirectional)
self.forward_start = [nn.VariableNode([r], None, name='forward_h_start'+str(i)) for i in range(self.config.p.gru_depth)]
self.vs += self.forward_start
if self.config.p.bidirectional:
self.backward_start = [nn.VariableNode([r], None, name='backward_h_start'+str(i)) for i in range(self.config.p.gru_depth)]
self.vs+=self.backward_start
else:
self.backward_start=None
out_size = r * self.config.p.gru_depth * (2 if self.config.p.bidirectional else 1)
self.main_first_W = nn.VariableNode([out_size, r], None, name='main_first_W')
self.main_first_b = nn.VariableNode([r], None, name='main_first_b')
self.main_Ws = [nn.VariableNode([r, r], None, name='main_W_'+str(i)) for i in range(self.config.p.out_layers)]
self.main_bs = [nn.VariableNode([r], None, name='main_b_'+str(i)) for i in range(self.config.p.out_layers)]
self.last_W = nn.VariableNode([r, 1], None, name='last_W_')
self.last_b = nn.VariableNode([1], None, name='last_b_')
self.vs+=[self.main_first_W, self.main_first_b,
self.last_W, self.last_b]+self.main_Ws + self.main_bs
self.rvs+=[self.main_first_W, self.last_W]+self.main_Ws
# we always need to add the trainer
self.add_trainer()
def __init__(self, config):
self.config = config
self.vs = []
self.rvs = []
r = self.config.p.r
num_tokens = self.config.num_tokens
# the embedding dictionary, which I think is the only shared variable
self.L = nn.VariableNode([num_tokens, self.config.p.r], None, name='L')
self.vs.append(self.L)
# add the attention matrix if needed
if self.config.p.attention and self.config.p.matrix_attention:
# add the attention matrix
left_size = (r * self.config.p.gru_depth
) if self.config.p.full_state_attention else r
# assume bidirectional
right_size = 2 * left_size if self.config.p.bidirectional else left_size
self.attention_B = nn.VariableNode([left_size, right_size],
None,
name='attention_B')
self.vs.append(self.attention_B)
def __init__(self, config):
DefaultVariables.__init__(self, config)
# define the specific variables for this section
r = self.config.p.r
out_layers = self.config.p.out_layers
# the parameters for the input section
self.known_gru_block = self.add_GRUb_block('known', bidirectional = self.config.p.bidirectional)
self.to_prove_gru_block = self.add_GRUb_block('to_prove', bidirectional = self.config.p.bidirectional)
self.out_gru_block = self.add_GRUb_block('out', bidirectional=False, takes_attention=self.config.p.attention)
self.forward_start = [nn.VariableNode([r], None, name='forward_h_start'+str(i)) for i in range(self.config.p.gru_depth)]
self.vs += self.forward_start
if self.config.p.bidirectional:
self.backward_start = [nn.VariableNode([r], None, name='backward_h_start'+str(i)) for i in range(self.config.p.gru_depth)]
self.vs+=self.backward_start
else:
self.backward_start=None
# the parameters for the output section
self.last_W = nn.VariableNode([r,self.config.num_tokens], None, name='last_W')
self.last_b = nn.VariableNode([self.config.num_tokens], None, name='last_b')
self.out_Ws = [nn.VariableNode([r,r], None, name='out_W_'+str(i)) for i in range(out_layers)]
self.out_bs = [nn.VariableNode([r], None, name='out_b_'+str(i)) for i in range(out_layers)]
self.vs += [self.last_W, self.last_b] + self.out_Ws + self.out_bs
self.rvs += [self.last_W] + self.out_Ws
# the parameters for the middle section
in_dim = 2*r if self.config.p.bidirectional else r
self.middle_W = [nn.VariableNode([in_dim, r], None, name='middle_W'+str(i)) for i in range(self.config.p.gru_depth)]
self.middle_b = [nn.VariableNode([r], None, name='middle_b'+str(i)) for i in range(self.config.p.gru_depth)]
self.vs += self.middle_W + self.middle_b
self.rvs += self.middle_W
# we always need to add the trainer
self.add_trainer()