Fix TFLite RESHAPE assert

[apivovarov]

Recently I found that Reshape op input might have 1 or 2 input tensors.
in TVM code we assert that length is 2 but we only use input_tensor[0].
Looks like the second input tensor was reshape shape in the past. But current TFLite schema which we use r1.13 has ReshapeOptions which contains new_shape:

table ReshapeOptions {
  new_shape:[int];
}

We do not use the second input tensor in TFLite frontend, we use ReshapeOptions instead.

What is more is that I have custom tflite model which has only one input tensor for RESHAPE op.
Current assert for input tensors length makes the compilation fail.
I suggest what we allow input tensors length for RESHAPE operator to be 1 or 2.

We can not change it to just 1 because test ops/models in test_forward.py have 2 input tensors for RESHAPE

[apivovarov]

@FrozenGene Can you have a look?

[FrozenGene]

could we add one unittest?

[u99127]

In addition I would request a comment in the code about the status in 1.13 and if possible the version of tflite when this changed.

[apivovarov]

@FrozenGene I tried to prepare tflite file with RESHAPE op having only one input tensor. But it always give me model with two input tensors for RESHAPE op.

Do you know how to force tflite converter to create RESHAPE op with only one tensor?

my models
https://www.dropbox.com/s/n8gg3gt2g3xlmkq/Screenshot%202019-11-13%2016.37.25.png?dl=0

https://www.dropbox.com/s/bhvuhrio277ka82/Screenshot%202019-11-13%2016.40.55.png?dl=0

hand_landmark.tflite model (has only one tensor in RESHAPE input)
https://www.dropbox.com/s/0wq2owq8p4hbw6l/Screenshot%202019-11-13%2016.37.49.png?dl=0

One of the solutions is to add hand_landmark.tflite to the end of tflite/test_forward.py. What you think? It will be good integration test for PRELU and RESHAPE.

[FrozenGene]

I can not construct it too. A bit strange.

We could add hand_landmark.tflite, however, we should also have unittest of reshape in my opinion.

[apivovarov]

TVM code does not use the second input tensor in RESHAPE op. From the beginning the assert should just check that input tensors array length is at least 1.
The reason for assert line is to display human readable error message instead of array index out of boundary exception. Because we only use input tensor with index 0 in the code the assert should check that input tensors array length is at least 1.
I'll add hand_landmark.tflite to the tests after your PR to fix PRELU is merged

[u99127]

I've been noticing differences in the tflite converter graph between 1.13 and 1.14 especially with split and unpack - have either of you guys played with multiple versions of the tfliteconverter ?

[apivovarov]

@FrozenGene All checks have passed - Can you approve?

[yzhliu]

Thanks @apivovarov @FrozenGene @u99127