[YOLO]yolo op added in frontend and removed from topi

[siju-samuel]

The yolo op is used only in darknet and since its not generic and can be acheived by frontend itself, i implemented the op in darknet.py and removed all the existing yolo_op registration, testcases and topi code.

@PariksheetPinjari909 @yzhliu Please review. Thanks in advance.

Thanks for contributing to TVM! Please refer to guideline https://docs.tvm.ai/contribute/ for useful information and tips. After the pull request is submitted, please request code reviews from Reviewers.

[siju-samuel]

cc @PariksheetPinjari909 @masahi @yzhliu please review

[PariksheetPinjari909]

As yolo, region are not generic ops and looking at the progress of darknet, moving these ops to frontend is a good idea. Thanks @siju-samuel for bringing in this PR. Let's check if we can bring region to frontend as well. Any views on this @yzhliu

[siju-samuel]

@PariksheetPinjari909 I have already tried this.
But im facing a strange issue with concatenate for some specific values in graph_runtime create.
I reported an issue here
https://discuss.tvm.ai/t/graphruntimecreate-is-in-deadloop/1032
Waiting for replys.
This can be replaced with the below code and all the region op in topi/nnvm can be removed.

def _darknet_region(inputs, attrs):
    """Process the region operation."""
    num = attrs.get('n', 1)
    classes = attrs.get('classes', 1)
    coords = attrs.get('coords', 0)
    background = attrs.get('background', 0)
    softmax = attrs.get('softmax', True)
    input_shape = attrs.get('shape')

    split_size = classes + coords + 1;
    intermediate_shape = (input_shape[0], num, split_size, input_shape[2], input_shape[3])
    data_block = _sym.reshape(inputs[0], shape=intermediate_shape)
    split_res = _sym.split(data_block, indices_or_sections=split_size, axis=2)
    splits = []
    for i in split_res:
        splits.append(i)
    splits[0] = _sym.sigmoid(splits[0])
    splits[1] = _sym.sigmoid(splits[1])
    if not background:
        splits[coords] = _sym.sigmoid(splits[coords])
    if softmax:
        offset = coords + int(not background)
        softmax_input = _sym.concatenate(*splits[offset:], axis=2)
        softmax_output = _sym.softmax(softmax_input, axis=2)
        data_block_1 = splits[:offset]
        splits = data_block_1 + [softmax_output]
    concat_out = _sym.concatenate(*splits, axis=2)
    return _sym.reshape(concat_out, shape=input_shape), None

[PariksheetPinjari909]

LGTM

[tqchen]

Thanks @PariksheetPinjari909 @siju-samuel this is merged