Method build_initial_coverage

caffe2/python/rnn_cell.py:1283–1326

initial_coverage is always zeros of shape [encoder_length], which shape must be determined programmatically dureing network computation. This method also sets self.coverage_weights, a separate transform of encoder_outputs which is used to determine coverage

(self, model)

        return self.decoder_cell.prepare_input(model, input_blob)

    def build_initial_coverage(self, model):
        """
        initial_coverage is always zeros of shape [encoder_length],
        which shape must be determined programmatically dureing network
        computation.

        This method also sets self.coverage_weights, a separate transform
        of encoder_outputs which is used to determine coverage contribution
        tp attention.
        """
        assert self.attention_type == AttentionType.SoftCoverage

        # [encoder_length, batch_size, encoder_output_dim]
        self.coverage_weights = brew.fc(
            model,
            self.encoder_outputs,
            self.scope('coverage_weights'),
            dim_in=self.encoder_output_dim,
            dim_out=self.encoder_output_dim,
            axis=2,
        )

        encoder_length = model.net.Slice(
            model.net.Shape(self.encoder_outputs),
            starts=[0],
            ends=[1],
        )
        if (
            scope.CurrentDeviceScope() is not None and
            core.IsGPUDeviceType(scope.CurrentDeviceScope().device_type)
        ):
            encoder_length = model.net.CopyGPUToCPU(
                encoder_length,
                'encoder_length_cpu',
            )
        # total attention weight applied across decoding steps_per_checkpoint
        # shape: [encoder_length]
        initial_coverage = model.net.ConstantFill(
            encoder_length,
            self.scope('initial_coverage'),
            value=0.0,
            input_as_shape=1,
        )
        return initial_coverage

    def get_state_names(self):
        state_names = list(self.decoder_cell.get_state_names())