def _create_tf_bot_embed(self, b_in: "tf.Tensor") -> "tf.Tensor":
"""Create embedding bot vector."""
b = train_utils.create_tf_fnn(
b_in,
self.hidden_layers_sizes["bot"],
self.droprate["bot"],
self.C2,
self._is_training,
layer_name_suffix="bot",
)
return train_utils.create_tf_embed(
b, self.embed_dim, self.C2, self.similarity_type, layer_name_suffix="bot"
)
def _create_tf_dial(self, a_in) -> Tuple["tf.Tensor", "tf.Tensor"]:
"""Create dialogue level embedding and mask."""
# mask different length sequences
# if there is at least one `-1` it should be masked
mask = tf.sign(tf.reduce_max(self.a_in, -1) + 1)
a = train_utils.create_tf_fnn(
a_in,
self.hidden_layers_sizes["pre_dial"],
self.droprate["dial"],
self.C2,
self._is_training,
layer_name_suffix="pre_dial",
)
self.attention_weights = {}
hparams = train_utils.create_t2t_hparams(
self.num_transformer_layers,
self.transformer_size,
self.num_heads,
self.droprate["dial"],
self.pos_encoding,
self.max_seq_length,
self._is_training,
)
a = train_utils.create_t2t_transformer_encoder(a, mask,
self.attention_weights,
hparams, self.C2,
self._is_training)
if isinstance(self.featurizer, MaxHistoryTrackerFeaturizer):
# pick last label if max history featurizer is used
a = a[:, -1:, :]
mask = mask[:, -1:]
dial_embed = train_utils.create_tf_embed(a,
self.embed_dim,
self.C2,
self.similarity_type,
layer_name_suffix="dial")
return dial_embed, mask
def _create_tf_embed_fnn(
self,
x_in: "tf.Tensor",
layer_sizes: List[int],
fnn_name: Text,
embed_name: Text,
) -> "tf.Tensor":
"""Create nn with hidden layers and name"""
x = train_utils.create_tf_fnn(
x_in,
layer_sizes,
self.droprate,
self.C2,
self._is_training,
layer_name_suffix=fnn_name,
)
return train_utils.create_tf_embed(
x,
self.embed_dim,
self.C2,
self.similarity_type,
layer_name_suffix=embed_name,
)
