[FRONTEND][TENSORFLOW] Simplify input shape

[alexeyr]

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.

Depends on #2105.

[alexeyr]

This gives some warnings when running tests, like

[14:16:31] /workspace/src/romanov/tvm/src/schedule/bound.cc:110: not in feed graph consumer = compute(input0_red_temp, 0x478efb0)
[14:16:31] /workspace/src/romanov/tvm/src/schedule/bound.cc:110: not in feed graph consumer = compute(input0_red_temp, 0x478efb0)

and

WARNING:autotvm:Cannot find config for target=llvm, workload=('conv2d', (4, 8, 8, 176, 'float32'), (1, 1, 176, 32, 'float32'), (1, 1), (0, 0), (1, 1), 'NHWC', 'float32'). A fallback configuration is used, which may bring great performance regression.

I guess the second should be fixed by modifying AutoTVM lookup, but I don't know about the first.

[icemelon]

@alexeyr Could you resolve the conflicts to latest master?

[srkreddy1238]

@alexeyr changes related to Unstack and Split(V) are already merged.
I think input_shape simplification may not be required as a node may have multiple outputs.

[alexeyr]

@icemelon9 Updated.
@srkreddy1238 This is taken into account: even if a node has multiple outputs, each input of another node uses only one of them and we still have a 1-element list there: https://github.com/dmlc/tvm/blob/021099f8d7d9cc59ae9ee872edde45803983d45c/nnvm/python/nnvm/frontend/tensorflow.py#L1235-L1243.

[alexeyr]

@srkreddy1238 Yes, this pull request doesn't fix any bugs, it is intended to simplify code while producing the exact same result. I've now verified that it actually does that:

I get the

...int_set.cc:514: cannot evaluate set type Cast

and other warnings when running the tests from IPython on the first try and don't get them after repeating. This happens both in master and with this pull request.

I also saved JSON of all graphs produced by test_forward.py and verified they are the same before and after the change.

Can you please review?

[srkreddy1238]

@alexeyr Thanks. LGTM.