Method _embed_gc

wavenet/model.py:531–566

Returns embedding for global condition. :param global_condition: Either ID of global condition for tf.nn.embedding_lookup or actual embedding. The latter is experimental. :return: Embedding or None

(self, global_condition)

        return encoded

    def _embed_gc(self, global_condition):
        '''Returns embedding for global condition.
        :param global_condition: Either ID of global condition for
               tf.nn.embedding_lookup or actual embedding. The latter is
               experimental.
        :return: Embedding or None
        '''
        embedding = None
        if self.global_condition_cardinality is not None:
            # Only lookup the embedding if the global condition is presented
            # as an integer of mutually-exclusive categories ...
            embedding_table = self.variables['embeddings']['gc_embedding']
            embedding = tf.nn.embedding_lookup(embedding_table,
                                               global_condition)
        elif global_condition is not None:
            # ... else the global_condition (if any) is already provided
            # as an embedding.

            # In this case, the number of global_embedding channels must be
            # equal to the the last dimension of the global_condition tensor.
            gc_batch_rank = len(global_condition.get_shape())
            dims_match = (global_condition.get_shape()[gc_batch_rank - 1] ==
                          self.global_condition_channels)
            if not dims_match:
                raise ValueError('Shape of global_condition {} does not'
                                 ' match global_condition_channels {}.'.
                                 format(global_condition.get_shape(),
                                        self.global_condition_channels))
            embedding = global_condition

        if embedding is not None:
            embedding = tf.reshape(
                embedding,
                [self.batch_size, 1, self.global_condition_channels])

        return embedding

    def predict_proba(self, waveform, global_condition=None, name='wavenet'):
        '''Computes the probability distribution of the next sample based on