Why in nn.conv2d, the padding can be a four dimension tuple?

haruhi · December 2, 2021, 11:36am

Before the relay.build, I print the Relay IR Module, but I find that:

nn.conv2d(%data, %graph_conv_weight, strides=[2, 2], padding=[1, 1, 1, 1], channels=32, kernel_size=[3, 3])

Can anyone help me to point out the direction in padding tuples? For example, padding[0] represents the height_top?

AndrewZhaoLuo · December 2, 2021, 6:26pm

apache/tvm/blob/f435a13a48b9545a9e5fa013ad0a0cc1e7c757d0/python/tvm/relay/op/nn/utils.py#L53


        else:
            raise ValueError("Size of padding can only be 1 or 2")
    elif isinstance(padding, int):
        pad_w = padding * 2
    else:
        raise ValueError("Unknown padding option %s" % padding)
    pad_left = (pad_w + 1) // 2
    return pad_left, pad_w - pad_left




def get_pad_tuple2d(padding):
    """Common code to get the pad option
    Parameters
    ----------
    padding : Union[int, Tuple[int, ...]]
        Padding size
    Returns
    -------
    pad_top : int
        Padding size on top
    pad_left : int

Appears to be [top, left, down, right]

haruhi · December 3, 2021, 1:09am

thank you very much@AndrewZhaoLuo