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

Before Change

            self.proj_key = nn.Dense(units=self._key_units, use_bias=self._use_bias,
                                     flatten=False, weight_initializer=weight_initializer,
                                     bias_initializer=bias_initializer, prefix="key_")
            self.proj_value = nn.Dense(units=self._value_units, use_bias=self._use_bias,
                                       flatten=False, weight_initializer=weight_initializer,
                                       bias_initializer=bias_initializer, prefix="value_")

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

After Change

        self._base_cell = base_cell
        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"

Instances