[Relay] Fix an adaptive_max_pool1d operator conversion bug

[haoyang9804]

This is a code patch for fixing this issue: #15004
Please refer to my comments in this link for the root cause of this issue.

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• cc @shingjan _{See #10317 for details}

_{Generated by tvm-bot}

[haoyang9804]

cc @jikechao

[jikechao]

@haoyang9804
LGTM, Thanks for your fixing.

[jikechao]

@echuraev @Hzfengsy @vvchernov
Please recheck it.

[Hzfengsy]

Please add a regression test for it

[vvchernov]

From one hand LGTM, from another one it confuses me due to fix for specific op was done on very high level (and CI test was constructed for specific op not for all cases). It looks like it has been assumed that outputs exist at any ret_name in ret_names. I need time to analyze it by my-self or explanation what is wrong with the initial assumption.

[haoyang9804]

From one hand LGTM, from another one it confuses me due to fix for specific op was done on very high level (and CI test was constructed for specific op not for all cases). It looks like it has been assumed that outputs exist at any ret_name in ret_names. I need time to analyze it by my-self or explanation what is wrong with the initial assumption.

I agree that "fix for specific op was done on very high level". So I make the fix only specific to the buggy pytorch op. This is only a workaround currently. I believe there are some deeper issues make this bug happen.

[jikechao]

cc @vvchernov

[echuraev]

LGTM. Thanks!