[Relay][TensorFlow] Add support for SquaredDifference

[soiferj]

Add support for SquaredDifference operator. Also include a minor bug fix for BatchMatMul.

@srkreddy1238 @alexeyr would you be able to take a look?

[alexeyr]

I don't have a problem with BatchMatMul change, but what is the bug it fixes? And should other attr['_output_shapes'] uses be changed too?

[soiferj]

I was testing out a BERT model which was exported with (add_shapes=True), but then ran the GraphTransform FoldConstants. For whatever reason, this seemed to remove shape information from the graph. In my testing, this was the only operator that threw this issue.

I don't think we need to do this for each operator quite yet. This fix is just a more future-proof way of coding this logic.

[yongwww]

I was testing out a BERT model which was exported with (add_shapes=True), but then ran the GraphTransform FoldConstants. For whatever reason, this seemed to remove shape information from the graph. In my testing, this was the only operator that threw this issue.

I don't think we need to do this for each operator quite yet. This fix is just a more future-proof way of coding this logic.

What happened if w/o (add_shapes=True)?

[soiferj]

Sorry, can you explain what you mean?

[yongwww]

@soiferj the tf converter works even without specifying add_shapes=True actually. So just wondering if there are any issues for the exported bert without explicitly specifying add_shapes=True.

[yongwww]

LGTM

[soiferj]

I see. I am not sure, I didn't try that out. However, we may want to think about not relaying on _output_shape in the future.

[yongwww]

@soiferj if your bert model is working, it will be nice to add an end to end test case with that model.

[soiferj]

Agreed. I will work on adding that in another PR once I’ve verified everything looks ok after this change.

[lixiaoquan]

I see. I am not sure, I didn't try that out. However, we may want to think about not relaying on _output_shape in the future.

It works, but it is very slow. Without shape information got from TF, Frontend will have to infer type for each new added node. For a model containing about 8000 nodes, it will take more than 1.5 hours to import a graph.

[tqchen]

@zhiics please help to manage this PR

[zhiics]

The change now looks good to me.

@alexeyr @yongwww @lixiaoquan please take another look and https://docs.tvm.ai/contribute/code_review.html#approve-and-request-changes-explicitly