def add_biaffine(self, inputs):
## inputs [seq_len, batch_size, units]
## first define four different MLPs
arc_roles = ['arc-dep', 'arc-head']
rel_roles = ['rel-dep', 'rel-head']
joint_roles = ['jk', 'stag']
vectors = {}
for arc_role in arc_roles:
for i in xrange(self.opts.mlp_num_layers):
if i == 0:
inputs_dim = self.outputs_dim
vector_mlp = inputs
else:
inputs_dim = self.opts.arc_mlp_units
weights = get_mlp_weights('{}_MLP_Layer{}'.format(arc_role, i), inputs_dim, self.opts.arc_mlp_units)
vector_mlp = self.add_dropout(tf.map_fn(lambda x: mlp(x, weights), vector_mlp), self.mlp_prob)
## [seq_len, batch_size, 2*mlp_units]
vectors[arc_role] = vector_mlp
weights = get_arc_weights('arc', self.opts.arc_mlp_units)
arc_output = arc_equation(vectors['arc-head'], vectors['arc-dep'], weights) # [batch_size, seq_len, seq_len] dim 1: deps, dim 2: heads
# arc_predictions = get_arcs(arc_output, self.test_opts) # [batch_size, seq_len]
arc_predictions = tf.argmax(arc_output, 2) # [batch_size, seq_len]
for rel_role in rel_roles:
for i in xrange(self.opts.mlp_num_layers):
if i == 0:
inputs_dim = self.outputs_dim
vector_mlp = inputs
else:
inputs_dim = self.opts.rel_mlp_units
weights = get_mlp_weights('{}_MLP_Layer{}'.format(rel_role, i), inputs_dim, self.opts.rel_mlp_units)
vector_mlp = self.add_dropout(tf.map_fn(lambda x: mlp(x, weights), vector_mlp), self.mlp_prob)
## [seq_len, batch_size, 2*mlp_units]
vectors[rel_role] = vector_mlp
weights = get_rel_weights('rel', self.opts.rel_mlp_units, self.loader.nb_rels)
rel_output, rel_scores = rel_equation(vectors['rel-head'], vectors['rel-dep'], weights, arc_predictions) #[batch_size, seq_len, nb_rels]
## joint stagging
for joint_role in joint_roles:
for i in xrange(self.opts.mlp_num_layers):
if i == 0:
inputs_dim = self.outputs_dim
vector_mlp = inputs
else:
inputs_dim = self.opts.joint_mlp_units
weights = get_mlp_weights('{}_MLP_Layer{}'.format(joint_role, i), inputs_dim, self.opts.joint_mlp_units)
vector_mlp = self.add_dropout(tf.map_fn(lambda x: mlp(x, weights), vector_mlp), self.mlp_prob)
## [seq_len, batch_size, 2*mlp_units]
vectors[joint_role] = vector_mlp
weights = get_joint_weights('stag', self.opts.joint_mlp_units, self.loader.nb_stags)
self.stag_embeddings = tf.transpose(weights['W-joint'], [1,0])
joint_output = joint_equation(vectors['stag'], weights) # [batch_size, seq_len, nb_stags]
weights = get_joint_weights('jk', self.opts.joint_mlp_units, self.loader.nb_jk)
joint_output_jk = joint_equation(vectors['jk'], weights) # [batch_size, seq_len, nb_stags]
return arc_output, rel_output, rel_scores, joint_output, joint_output_jk
def add_biaffine_global(self, inputs):
## inputs [seq_len, batch_size, units] = [n, b, d]
## first define four different MLPs
arc_roles = ['arc-dep', 'arc-head']
rel_roles = ['rel-dep', 'rel-head']
vectors = {}
for arc_role in arc_roles:
for i in xrange(self.opts.mlp_num_layers):
if i == 0:
inputs_dim = self.outputs_dim
vector_mlp = inputs
else:
inputs_dim = self.opts.arc_mlp_units
weights = get_mlp_weights('{}_MLP_Layer{}'.format(arc_role, i),
inputs_dim, self.opts.arc_mlp_units)
vector_mlp = self.add_dropout(
tf.map_fn(lambda x: mlp(x, weights), vector_mlp),
self.mlp_prob)
vectors[arc_role] = vector_mlp
## [seq_len, batch_size, arc_mlp_units]
arc_weights = get_arc_weights('arc', self.opts.arc_mlp_units)
for rel_role in rel_roles:
for i in xrange(self.opts.mlp_num_layers):
if i == 0:
inputs_dim = self.outputs_dim
vector_mlp = inputs
else:
inputs_dim = self.opts.rel_mlp_units
weights = get_mlp_weights('{}_MLP_Layer{}'.format(rel_role, i),
inputs_dim, self.opts.rel_mlp_units)
vector_mlp = self.add_dropout(
tf.map_fn(lambda x: mlp(x, weights), vector_mlp),
self.mlp_prob)
vectors[rel_role] = vector_mlp
## [seq_len, batch_size, rel_mlp_units]
rel_weights = get_rel_weights('rel', self.opts.rel_mlp_units,
self.loader.nb_rels)
global_weights = get_global_weights('global', self.opts.arc_mlp_units,
self.loader.nb_rels)
arc_output, rel_output, rel_scores = global_equation(
vectors, arc_weights, rel_weights, global_weights)
return arc_output, rel_output, rel_scores