def __init__(self,
model: dy.ParameterCollection,
in_dim: int,
hid_dim: int,
p: int = 0.1):
pc = model.add_subcollection()
self.W1 = Linear(pc, in_dim, hid_dim)
self.W2 = Linear(pc, hid_dim, in_dim)
self.p = p
self.pc = pc
self.spec = (in_dim, hid_dim, p)
def __init__(self,
model: dy.ParameterCollection,
in_dim: int,
out_dim: int,
init: dy.PyInitializer = None,
bias: bool = True):
pc = model.add_subcollection()
if not init: init = dy.UniformInitializer(math.sqrt(in_dim))
self.W = pc.add_parameters((out_dim, in_dim), init=init)
if bias: self.b = pc.add_parameters((out_dim, ), init=init)
self.pc = pc
self.bias = bias
def __init__(self, model: dy.ParameterCollection, cfg: IniConfigurator,
vocabulary: Vocabulary):
pc = model.add_subcollection()
# MLP layer
orth_init = OrthogonalInitializer
self.head_arc_MLP = MLP(pc, cfg.ARC_MLP_SIZE, leaky_relu, cfg.MLP_DROP,
cfg.MLP_BIAS, orth_init)
self.head_rel_MLP = MLP(pc, cfg.REL_MLP_SIZE, leaky_relu, cfg.MLP_DROP,
cfg.MLP_BIAS, orth_init)
self.dept_arc_MLP = MLP(pc, cfg.ARC_MLP_SIZE, leaky_relu, cfg.MLP_DROP,
cfg.MLP_BIAS, orth_init)
self.dept_rel_MLP = MLP(pc, cfg.REL_MLP_SIZE, leaky_relu, cfg.MLP_DROP,
cfg.MLP_BIAS, orth_init)
# Biaffine Attention Layer (Arc)
arc_size = cfg.ARC_MLP_SIZE[-1]
zero_init = dy.ConstInitializer(0)
self.arc_attn_mat = [
BiaffineMatAttention(pc, arc_size, arc_size, 1, True, False,
zero_init)
for _ in range(cfg.GRAPH_LAYERS + 1)
]
# Biaffine Attention Layer (Rel)
rel_num = vocabulary.get_vocab_size('rel')
rel_size = cfg.REL_MLP_SIZE[-1]
self.rel_mask = np.array([1] + [0] *
(rel_num - 1)) # mask root relation
self.rel_attn = BiaffineMatAttention(pc, rel_size, rel_size, rel_num,
True, True, zero_init)
# Graph Network Layer
self.head_gnn = GraphNNUnit(pc, arc_size, arc_size, leaky_relu,
orth_init)
self.dept_gnn = GraphNNUnit(pc, arc_size, arc_size, leaky_relu,
orth_init)
self.head_rel_gnn = GraphNNUnit(pc, rel_size, rel_size, leaky_relu,
orth_init)
self.dept_rel_gnn = GraphNNUnit(pc, rel_size, rel_size, leaky_relu,
orth_init)
# Graph Layer WarmUp
self.warm_list = [
-i * cfg.WARM for i in range(cfg.GRAPH_LAYERS, -1, -1)
]
# Save Variable
self.arc_size, self.rel_size, self.rel_num = arc_size, rel_size, rel_num
self.pc, self.cfg = pc, cfg
self.spec = (cfg, vocabulary)
def __init__(self,
model: dy.ParameterCollection,
in_dim: int,
out_dim: int,
bias: bool = True,
init: dy.PyInitializer = dy.GlorotInitializer()):
pc = model.add_subcollection()
init = init_wrap(init, (out_dim, in_dim))
self.W = pc.add_parameters((out_dim, in_dim), init=init)
if bias:
self.b = pc.add_parameters((out_dim, ), init=0)
self.pc = pc
self.bias = bias
self.spec = (in_dim, out_dim, bias, init)
def __init__(
self,
model: dy.ParameterCollection,
h_dim: int,
d_dim: int,
f=dy.tanh,
init: dy.PyInitializer = dy.GlorotInitializer()):
pc = model.add_subcollection()
init_W = init_wrap(init, (h_dim, d_dim))
self.W = pc.add_parameters((h_dim, d_dim), init=init_W)
init_B = init_wrap(init, (h_dim, h_dim))
self.B = pc.add_parameters((h_dim, h_dim), init=init_B)
self.pc, self.f = pc, f
self.spec = (h_dim, d_dim, f, init)
def __init__(self,
model: dy.ParameterCollection,
sizes: List[int],
f: 'nonlinear' = dy.tanh,
p: float = 0.0,
bias: bool = True,
init: dy.PyInitializer = dy.GlorotInitializer()):
pc = model.add_subcollection()
self.W = [
pc.add_parameters((x, y), init=init_wrap(init, (x, y)))
for x, y in zip(sizes[1:], sizes[:-1])
]
if bias:
self.b = [pc.add_parameters((y, ), init=0) for y in sizes[1:]]
self.pc, self.f, self.p, self.bias = pc, f, p, bias
self.spec = (sizes, f, p, bias, init)