Method __init__

python/lbann/modules/base.py:230–341

Initialize convolution module. Args: num_dims (int): Number of dimensions. out_channels (int): Number of output channels, i.e. number of filters. kernel_size (int) or (list): Size of convolution kernel. Either an int for square kernel or l

(self,
                 num_dims,
                 out_channels,
                 kernel_size,
                 stride=1,
                 padding=0,
                 dilation=1,
                 groups=1,
                 bias=True,
                 weights=[],
                 activation=None,
                 name=None,
                 transpose=False,
                 parallel_strategy={})

    global_count = 0  # Static counter, used for default names

    def __init__(self,
                 num_dims,
                 out_channels,
                 kernel_size,
                 stride=1,
                 padding=0,
                 dilation=1,
                 groups=1,
                 bias=True,
                 weights=[],
                 activation=None,
                 name=None,
                 transpose=False,
                 parallel_strategy={}):
        """Initialize convolution module.

        Args:
            num_dims (int): Number of dimensions.
            out_channels (int): Number of output channels, i.e. number
                of filters.
            kernel_size (int) or (list): Size of convolution kernel. Either an int for square kernel or list of size num_dims.
            has_vector (bool): If true then call with non-square kernel
              padding, stride, dilation, and padding
            stride (int) or (list): Convolution stride. Either an int for square kernel or list of size num_dims.
            padding (int) or (list): Convolution padding. Either an int for square kernel or list of size num_dims.
            dilation (int) or (list): Convolution dilation. Either an int for square kernel or list of size num_dims.
            groups (int): Number of convolution groups.
            bias (bool): Whether to apply channel-wise bias after
                convolution.
            weights (`Weights` or iterator of `Weights`): Weights in
                convolution layer. There are at most two: the kernel
                and the bias. If weights are not provided, the kernel
                will be initialized with He normal initialization and
                the bias with zeros.
            name (str): Default name is in the form 'convmodule<index>'.
            transpose (bool): If true call deconvolution (or convolution
                         transpose)
            parallel_strategy dict): Data partitioning scheme.

        """
        super().__init__()
        ConvolutionModule.global_count += 1
        self.name = (name
                     if name
                     else 'convmodule{0}'.format(ConvolutionModule.global_count))

        self.instance = 0
        self.num_dims = num_dims
        self.out_channels = out_channels

        self.kernel_dims = list(make_iterable(kernel_size))

        if (len(self.kernel_dims)) == 1:
          self.kernel_dims = self.kernel_dims * self.num_dims
        elif (len(self.kernel_dims)) != self.num_dims:
          raise ValueError("Invalid kernel dimensions passed to {}".format(self.name))

        self.stride = list(make_iterable(stride))