def __init__(self, conf: ZStackingLabelConf, base_node: ZlabelNode,
**kwargs):
# --
conf.__dict__.update(
base_node.conf.__dict__) # update from the base one!
# --
conf.emb_size = -1 # no use of basic ones!
super().__init__(conf, **kwargs)
conf: ZStackingLabelConf = self.conf
# --
self.setattr_borrow("_base", base_node)
#
assert conf.use_input_emb or conf.use_input_base
dim_emb = base_node.conf.emb_size
dim_base = base_node.conf._csize
if conf.use_input_emb:
self.aff_final = AffineNode(None,
isize=dim_emb,
osize=conf._csize,
no_drop=True)
else:
self.aff_final = None
if conf.use_input_base:
self.W_base = BK.new_param([dim_base, conf._csize])
BK.init_param(self.W_base, "glorot")
else:
self.W_base = None
def reset_parameters(self):
BK.init_param(self.mixed_weights, "zero") # make it all 0
if self.input_mask is not None:
with BK.no_grad_env():
self.input_mask.set_(
BK.random_bernoulli((self.conf._isize, ),
self.fixed_mask_rate, 1.))
def reset_parameters(self):
_isize = self.conf._isize
for ii in range(3):
BK.init_param(self.W_hid[:, ii * _isize:(ii + 1) * _isize],
"glorot")
BK.init_param(self.W_gate, "glorot")
BK.init_param(self.b_hid, "zero")
BK.init_param(self.b_gate, "zero")
def reset_parameters(self):
if self.use_lrank:
BK.init_param(self.E, "glorot", lookup=True)
BK.init_param(self.W, "ortho")
else:
BK.init_param(self.W, "zero")
def reset_parameters(self):
BK.init_param(self.w_out,
"glorot",
lookup=True,
scale=self.conf.init_scale_out)
def reset_parameters(self):
if self.logs2 is not None:
BK.init_param(self.logs2, "zero")
def reset_parameters(self):
if not self.conf.e_tie_weights:
BK.init_param(self.W, "glorot", lookup=True)