def __init__(self, hparams=None):
EncoderBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
if self._hparams.initializer:
tf.get_variable_scope().set_initializer(
layers.get_initializer(self._hparams.initializer))
self.multihead_attention_list = []
self.poswise_networks = []
for i in range(self._hparams.num_blocks):
with tf.variable_scope("layer_{}".format(i)):
with tf.variable_scope('attention'):
mh_attn = MultiheadAttentionEncoder(
self._hparams.multihead_attention)
self.multihead_attention_list.append(mh_attn)
if self._hparams.dim != mh_attn.hparams.output_dim:
raise ValueError(
'The "dim" in the hparams of '
'"multihead_attention" should be equal to the '
'"dim" of TransformerEncoder')
pw_net = FeedForwardNetwork(
hparams=self._hparams['poswise_feedforward'])
final_dim = pw_net.hparams.layers[-1]['kwargs']['units']
if self._hparams.dim != final_dim:
raise ValueError(
'The output dimenstion of '
'"poswise_feedforward" should be equal '
'to the "dim" of TransformerEncoder.')
self.poswise_networks.append(pw_net)
def __init__(self, vocab_size=None, output_layer=None, hparams=None):
ModuleBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
if self._hparams.initializer:
tf.get_variable_scope().set_initializer(
layers.get_initializer(self._hparams.initializer))
# Make the output layer
self._output_layer, self._vocab_size = _make_output_layer(
output_layer, vocab_size, self._hparams.output_layer_bias,
self.variable_scope)
# Make attention and poswise networks
self.multihead_attentions = {'self_att': [], 'encdec_att': []}
self.poswise_networks = []
for i in range(self._hparams.num_blocks):
layer_name = 'layer_{}'.format(i)
with tf.variable_scope(layer_name):
with tf.variable_scope("self_attention"):
multihead_attention = MultiheadAttentionEncoder(
self._hparams.multihead_attention)
self.multihead_attentions['self_att'].append(
multihead_attention)
if self._hparams.dim != \
multihead_attention.hparams.output_dim:
raise ValueError('The output dimenstion of '
'MultiheadEncoder should be equal '
'to the dim of TransformerDecoder')
with tf.variable_scope('encdec_attention'):
multihead_attention = MultiheadAttentionEncoder(
self._hparams.multihead_attention)
self.multihead_attentions['encdec_att'].append(
multihead_attention)
if self._hparams.dim != \
multihead_attention.hparams.output_dim:
raise ValueError('The output dimenstion of '
'MultiheadEncoder should be equal '
'to the dim of TransformerDecoder')
pw_net = FeedForwardNetwork(
hparams=self._hparams['poswise_feedforward'])
final_dim = pw_net.hparams.layers[-1]['kwargs']['units']
if self._hparams.dim != final_dim:
raise ValueError(
'The output dimenstion of '
'"poswise_feedforward" should be equal '
'to the "dim" of TransformerDecoder.')
self.poswise_networks.append(pw_net)
# Built in _build()
self.context = None
self.context_sequence_length = None
self.embedding = None
self._helper = None
self._cache = None
self.max_decoding_length = None
def __init__(self,
vocab_size=None,
output_layer=None,
tau=None,
hparams=None):
EncoderBase.__init__(self, hparams)
with tf.variable_scope(self.variable_scope):
if self._hparams.initializer:
tf.get_variable_scope().set_initializer(
layers.get_initializer(self._hparams.initializer))
# Make the output layer
self._output_layer, self._vocab_size = _make_output_layer(
output_layer, vocab_size, self._hparams.output_layer_bias,
self.variable_scope)
# Make attention and poswise networks
self.graph_multihead_attention_list = []
self.poswise_networks = []
for i in range(self._hparams.num_blocks):
with tf.variable_scope("layer_{}".format(i)):
with tf.variable_scope('attention'):
mh_attn = GraphMultiheadAttentionEncoder(
self._hparams.graph_multihead_attention)
self.graph_multihead_attention_list.append(mh_attn)
if self._hparams.dim != mh_attn.hparams.output_dim:
raise ValueError(
'The "dim" in the hparams of '
'"multihead_attention" should be equal to the '
'"dim" of CrossGraphTransformerFixedLengthDecoder'
)
pw_net = FeedForwardNetwork(
hparams=self._hparams['poswise_feedforward'])
final_dim = pw_net.hparams.layers[-1]['kwargs']['units']
if self._hparams.dim != final_dim:
raise ValueError(
'The output dimenstion of '
'"poswise_feedforward" should be equal '
'to the "dim" of CrossGraphTransformerFixedLengthDecoder.'
)
self.poswise_networks.append(pw_net)
self._helper = None
self._tau = tau
def test_feedforward(self):
"""Tests feed-forward.
"""
hparams = {
"layers": [{
"type": "Dense",
}, {
"type": "Dense",
}]
}
nn = FeedForwardNetwork(hparams=hparams)
self.assertEqual(len(nn.layers), len(hparams["layers"]))
_ = nn(tf.ones([64, 16, 16]))
self.assertEqual(len(nn.trainable_variables),
len(hparams["layers"]) * 2)
self.assertEqual(len(nn.layer_outputs), len(hparams["layers"]))