[Frontend][TENSORFLOW] batch_norm() got an unexpected keyword argument 'exponential_avg_factor'

[Leslie-Fang]

I am trying to convert a resnet style model into TVM, but I got the error message:

Traceback (most recent call last):
  File "tensorflow_tvm.py", line 47, in <module>
    mod, params = relay.frontend.from_tensorflow(graph_def,
  File "/home/lesliefang/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 3574, in from_tensorflow
    mod, params = g.from_tensorflow(graph, layout, shape, outputs)
  File "/home/lesliefang/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 2967, in from_tensorflow
    func = self._get_relay_func(graph, layout=layout, shape=shape, outputs=outputs)
  File "/home/lesliefang/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 2926, in _get_relay_func
    self._backtrack_construct(node.name)
  File "/home/lesliefang/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 3508, in _backtrack_construct
    op = self._convert_operator(node.op, inputs, attr, self._graph)
  File "/home/lesliefang/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 3365, in _convert_operator
    sym = convert_map[op_name](inputs, attrs, self._params, self._mod)
  File "/home/lesliefang/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 1265, in _impl
    out = AttrCvt(op_name='batch_norm',
  File "/home/lesliefang/tvm/tvm/python/tvm/relay/frontend/common.py", line 417, in __call__
    return get_relay_op(op_name)(*inputs, **new_attrs)
TypeError: batch_norm() got an unexpected keyword argument 'exponential_avg_factor'

Refer to this issue #3428, I have checked my model's bn op. it doesn't set this attr. But since this PR tensorflow/tensorflow#37176, tensorflow should set the default value of this attr.

I have 2 questions:

1. It seems TVM will also check the op definition in Tensorflow's implementation, why not just check the user-defined attribution in the PB model?
2. is it ok for me to just add the

self._ignores.append('exponential_avg_factor')

here? https://github.com/apache/incubator-tvm/blob/512ed3930a61daf38e80e1f71e51f0d1f139fb8e/python/tvm/relay/frontend/common.py#L367-L374

[Leslie-Fang]

@FrozenGene Could you help to take a look of this issue:) Thanks

[tqchen]

Thanks for reporting the problem, the community uses https://discuss.tvm.ai/ for quick trouble shooting and discussions, please open a new thread there

[mikeseven]

@tqchen I don't see the issue solved with top of tree today.
but adding @Leslie-Fang solution to common.py does work. I think it's ok because these ops are training ops that should have been removed from frozen graphs....
Same thing for RandomUniform, FIFOQueue(V2) , QueueDequeueMany(V2)?