Function MultiTaskProjection

Applies a multi-task projection. Calculates the projection of a batched input tensor, where the projection parameters (weights, bias) can be from a different 'task' for each batch. Args: weights: projection weight matrices of all tasks, size [num_tasks, input_dim, output_dim] b

(
    weights: tf.Tensor,
    biases: Optional[tf.Tensor],
    inputs: tf.Tensor,
    tasks: tf.Tensor,
    einsum_order: str,
    quant_layer,  # quant_utils.QuantizableLayer, would be circular import
    w_q_name: str,
    w_q_domain: str = 'default',
    qat_output: bool = False,
)