Function _get_same_padding_conv_nd

monai/networks/nets/efficientnet.py:808–847

Helper for getting padding (nn.ConstantPadNd) to be used to get SAME padding conv operations similar to Tensorflow's SAME padding. This function is generalized for MONAI's N-Dimensional spatial operations (e.g. Conv1D, Conv2D, Conv3D) Args: image_size: input image/feature

(
    image_size: list[int], kernel_size: tuple[int, ...], dilation: tuple[int, ...], stride: tuple[int, ...]
)

def _get_same_padding_conv_nd(
    image_size: list[int], kernel_size: tuple[int, ...], dilation: tuple[int, ...], stride: tuple[int, ...]
) -> list[int]:
    """
    Helper for getting padding (nn.ConstantPadNd) to be used to get SAME padding
    conv operations similar to Tensorflow's SAME padding.

    This function is generalized for MONAI's N-Dimensional spatial operations (e.g. Conv1D, Conv2D, Conv3D)

    Args:
        image_size: input image/feature spatial size.
        kernel_size: conv kernel's spatial size.
        dilation: conv dilation rate for Atrous conv.
        stride: stride for conv operation.

    Returns:
        paddings for ConstantPadNd padder to be used on input tensor to conv op.
    """
    # get number of spatial dimensions, corresponds to kernel size length
    num_dims = len(kernel_size)

    # additional checks to populate dilation and stride (in case they are single entry tuples)
    if len(dilation) == 1:
        dilation = dilation * num_dims

    if len(stride) == 1:
        stride = stride * num_dims

    # equation to calculate (pad^+ + pad^-) size
    _pad_size: list[int] = [
        max((math.ceil(_i_s / _s) - 1) * _s + (_k_s - 1) * _d + 1 - _i_s, 0)
        for _i_s, _k_s, _d, _s in zip(image_size, kernel_size, dilation, stride)
    ]
    # distribute paddings into pad^+ and pad^- following Tensorflow's same padding strategy
    _paddings: list[tuple[int, int]] = [(_p // 2, _p - _p // 2) for _p in _pad_size]

    # unroll list of tuples to tuples, and then to list
    # reversed as nn.ConstantPadNd expects paddings starting with last dimension
    _paddings_ret: list[int] = [outer for inner in reversed(_paddings) for outer in inner]
    return _paddings_ret


def _make_same_padder(conv_op: nn.Conv1d | nn.Conv2d | nn.Conv3d, image_size: list[int]):