Fix quantize pass error when excluding a quantization supported op

[ciyongch]

Description

This PR is to address the error when excluding a quantization supported op in a model like the case below, where both op1 and op2 support quantization.
op1 -> op2
When excluding op1 from the model, the pattern is changed to
op1 -> _contrib_quantize -> quantized_op2
There's an incorrect connection (incorrect entry index) between _contrib_quantize and quantize_op2 for min/max inputs in current code base.

Checklist

Essentials

Please feel free to remove inapplicable items for your PR.

• The PR title starts with [MXNET-$JIRA_ID], where $JIRA_ID refers to the relevant JIRA issue created (except PRs with tiny changes)
• Changes are complete (i.e. I finished coding on this PR)
• All changes have test coverage:
• Unit tests are added for small changes to verify correctness (e.g. adding a new operator)
• Nightly tests are added for complicated/long-running ones (e.g. changing distributed kvstore)
• Build tests will be added for build configuration changes (e.g. adding a new build option with NCCL)
• Code is well-documented:
• For user-facing API changes, API doc string has been updated.
• For new C++ functions in header files, their functionalities and arguments are documented.
• For new examples, README.md is added to explain the what the example does, the source of the dataset, expected performance on test set and reference to the original paper if applicable
• Check the API doc at http:https://mxnet-ci-doc.s3-accelerate.dualstack.amazonaws.com/PR-$PR_ID/$BUILD_ID/index.html
• To the my best knowledge, examples are either not affected by this change, or have been fixed to be compatible with this change

Changes

• Feature1, tests, (and when applicable, API doc)
• Feature2, tests, (and when applicable, API doc)

Comments

• If this change is a backward incompatible change, why must this change be made.
• Interesting edge cases to note here

@TaoLv @pengzhao-intel @apeforest

[pengzhao-intel]

Thanks for the fix.
Could you attach the picture of the graph before and after the change?

[ciyongch]

@pengzhao-intel The picture of graph might not reflect the error, an enhanced test was added to cover this kind of cases. I paste the full symbol file of the test case w/ and w/o the fix.

• symbol file w/o fix exclude_pool0.txt
• symbol file w/ fix exclude_pool0_fix.txt

The incorrect connection (error entry index) is located at 3nd/4th input (counting from 0), [21, 4294967295, 0], [21, 0, 0]

    {
      "op": "_contrib_quantized_conv", 
      "name": "quantized_conv1", 
      "attrs": {
        "kernel": "(1, 1)", 
        "no_bias": "False", 
        "num_filter": "4", 
        "pad": "(0, 0)"
      }, 
      "inputs": [[21, 0, 0], [22, 0, 0], [23, 0, 0], [21, 4294967295, 0], [21, 0, 0], [24, 0, 0], [25, 0, 0], [26, 0, 0], [27, 0, 0]]
    },

After fixed, the connection is as below, 3nd/4th inputs are correct now [21, 1, 0], [21, 2, 0]

    {
      "op": "_contrib_quantized_conv", 
      "name": "quantized_conv1", 
      "attrs": {
        "kernel": "(1, 1)", 
        "no_bias": "False", 
        "num_filter": "4", 
        "pad": "(0, 0)"
      }, 
      "inputs": [[21, 0, 0], [22, 0, 0], [23, 0, 0], [21, 1, 0], [21, 2, 0], [24, 0, 0], [25, 0, 0], [26, 0, 0], [27, 0, 0]]
    }, 
    {

[pengzhao-intel]

Thanks for the explanation. It makes sense.

[ciyongch]

ping @apeforest @TaoLv to have a reveiw :)

[TaoLv]

LGTM. Thank you for the fix.

[roywei]

@mxnet-label-bot add [Operator, pr-awaiting-review]
@ciyongch Thanks for the contribution, is there an issue related to this fix?

[ciyongch]

@roywei No open issue related to it so far. We found this error when trying to quantize Resnet50_v1 at local (with excluding some ops which supports quantization from the model script manually). In this case, NeedQuantize(e.node, excluded_nodes) is much more accurate than quantized_op_map.count(e.node->op()).
A test case is also added to cover this kind of error.