Aten::lt, aten::gt are not supported in tvm pytorch frontend

pyjhzwh · March 5, 2020, 4:28pm

I try to implement self-defined networks which involves comparing outputs to some thresholds. But it raises NotImplementedError: The following operators are not implemented: ['aten::gt', 'aten::lt', 'aten::ge']

I attempted to add functions to support such operators, like

github.com

apache/incubator-tvm/blob/f63b249dd01b8b59258a1ff310e957c5b1d16dd9/python/tvm/relay/frontend/pytorch.py#L579


            end[axis] = dim
            stride = [1] * len(shape)


            chunk_out = _op.transform.strided_slice(data, begin, end, stride)
            chunks.append(chunk_out)


        return chunks
    return _impl


def _matmul():
    def _impl(inputs, input_types):
        data0 = inputs[0]
        data1 = inputs[1]
        data1_t = _op.transpose(data1, axes=(1, 0))


        return _op.nn.dense(data0, data1_t)
    return _impl


def _expand():
    def _impl(inputs, input_types):
        data_in = inputs[0]

But I don’t know what are the elements in inputs of aten::lt or aten::gt stand for. And I wonder if there are exsiting functions in _op supported lt, gt, le, ge operations. Could you give some hints. Thanks.

pyjhzwh · March 5, 2020, 5:58pm

"aten::lt"                              : _elemwise("less"),
"aten::gt"                              : _elemwise("greater"),
"aten::le"                              : _elemwise("less_equal"),
"aten::ge"                              : _elemwise("greater_equal"),

Probably add codes above in _convert_map{}. But I not sure if it is correct

masahi · March 5, 2020, 10:16pm

some of them are in my ongoing pr https://github.com/apache/incubator-tvm/pull/4964

UPDATE: Added more operators in that PR. All of you requested are in.