b1b61d3f90cf795c7b48b6d109db7b7b96fa21ff,src/gluonnlp/model/attention_cell.py,MultiHeadAttentionCell,__init__,#MultiHeadAttentionCell#Any#Any#Any#Any#Any#Any#Any#Any#Any#Any#,206

Before Change

        self._value_units = value_units
        self._num_heads = num_heads
        self._use_bias = use_bias
        if self._query_units % self._num_heads != 0:
            raise ValueError("In MultiHeadAttetion, the query_units should be divided exactly"
                             " by the number of heads. Received query_units={}, num_heads={}"
                             .format(key_units, num_heads))

        if self._key_units % self._num_heads != 0:
            raise ValueError("In MultiHeadAttetion, the key_units should be divided exactly"
                             " by the number of heads. Received key_units={}, num_heads={}"
                             .format(key_units, num_heads))

        if self._value_units % self._num_heads != 0:
            raise ValueError("In MultiHeadAttetion, the value_units should be divided exactly"
                             " by the number of heads. Received value_units={}, num_heads={}"
                             .format(value_units, num_heads))

After Change

        self._num_heads = num_heads
        self._use_bias = use_bias
        units = {"query": query_units, "key": key_units, "value": value_units}
        for name, unit in units.items():
            if unit % self._num_heads != 0:
                raise ValueError(
                    "In MultiHeadAttetion, the {name}_units should be divided exactly"
                    " by the number of heads. Received {name}_units={unit}, num_heads={n}".format(
                        name=name, unit=unit, n=num_heads))
            setattr(self, "_{}_units".format(name), unit)
            with self.name_scope():
                setattr(
                    self, "proj_{}".format(name),
                    nn.Dense(units=self._query_units, use_bias=self._use_bias, flatten=False,
                             weight_initializer=weight_initializer,
                             bias_initializer=bias_initializer, prefix="{}_".format(name)))

    def __call__(self, query, key, value=None, mask=None):
        Compute the attention.

        Parameters

Instances