Class FullyConnectedModule

python/lbann/modules/base.py:38–132

Basic block for fully-connected neural networks. Applies a dense linearity and a nonlinear activation function.

        return self.forward(*args, **kwargs)

class FullyConnectedModule(Module):
    """Basic block for fully-connected neural networks.

    Applies a dense linearity and a nonlinear activation function.

    """

    global_count = 0  # Static counter, used for default names

    def __init__(self,
                 size,
                 bias=True,
                 transpose=False,
                 weights=[],
                 activation=None,
                 name=None,
                 data_layout='data_parallel',
                 parallel_strategy={}):
        """Initialize fully-connected module.

        Args:
            size (int): Size of output tensor.
            activation (type): Layer class for activation function.
            bias (bool): Whether to apply bias after linearity.
            transpose (bool): Whether to apply transpose of weights
                matrix.
            weights (`Weights` or iterator of `Weights`): Weights in
                fully-connected layer. There are at most two: the
                matrix and the bias. If weights are not provided, the
                matrix will be initialized with He normal
                initialization and the bias with zeros.
            name (str): Default name is in the form 'fcmodule<index>'.
            data_layout (str): Data layout.
            parallel_strategy (dict): Data partitioning scheme.

        """
        super().__init__()
        FullyConnectedModule.global_count += 1
        self.instance = 0
        self.size = size
        self.bias = bias
        self.transpose = transpose
        self.data_layout = data_layout
        self.parallel_strategy = parallel_strategy

        self.name = (name
                     if name
                     else 'fcmodule{0}'.format(FullyConnectedModule.global_count))

        # Initialize weights
        # Note: If weights are not provided, matrix weights are
        # initialized with He normal scheme and bias weights are
        # initialized with zeros.
        self.weights = list(make_iterable(weights))
        if len(self.weights) > 2:
            raise ValueError('`FullyConnectedModule` has '
                             'at most two weights, '
                             'but got {0}'.format(len(self.weights)))