[TFLite] Add transpose_conv to TFLite parser

[apivovarov]

This PR adds transpose_conv to TFLite frontend.

conv2d_transpose TF docs:
https://www.tensorflow.org/versions/r1.13/api_docs/python/tf/nn/conv2d_transpose

TFLite schema

table TransposeConvOptions {
  padding:Padding;
  stride_w:int;
  stride_h:int;
}

Limitations:

• Similar to TF frontend TFLite frontend supports 'SAME' padding for kernel 1x1 only.

[apivovarov]

@FrozenGene Could you have a look?

[FrozenGene]

if kh kw is 3x3, what is the current error msg?

[apivovarov]

Where? in TF or TFLite?

[apivovarov]

in TFLite it is

  File "test_forward.py", line 519, in test_forward_transpose_conv
    _test_transpose_conv([4, 32, 32, 16], [3, 3, 5, 16], [4, 32, 32, 5], [1, 1], 'SAME')

  File "test_forward.py", line 504, in _test_transpose_conv
    compare_tflite_with_tvm(data_array, 'Placeholder:0', [in_data], [out])

  File "test_forward.py", line 162, in compare_tflite_with_tvm
    num_output=len(out_names), out_names=out_names)

  File "test_forward.py", line 75, in run_tvm_graph
    dtype_dict=dtype_dict)

  File "/home/dlc/workplace/apivovarov/incubator-tvm/python/tvm/relay/frontend/tflite.py", line 1572, in from_tflite
    op_converter.convert_op_to_relay()

  File "/home/dlc/workplace/apivovarov/incubator-tvm/python/tvm/relay/frontend/tflite.py", line 125, in convert_op_to_relay
    ret = self.convert_map[op_code_str](op)

  File "/home/dlc/workplace/apivovarov/incubator-tvm/python/tvm/relay/frontend/tflite.py", line 1440, in convert_transpose_conv
    "SAME padding is supported for kernel (1,1) only"

AssertionError: SAME padding is supported for kernel (1,1) only

[FrozenGene]

Both should be the same. i.e. when our model is 3x3 conv_transpose, what is the current error msg of TVM?

[FrozenGene]

I mean if we remove this assert, what will be happened of TVM? If we add this assert, we can not support 3x3 conv_transpose

[apivovarov]

in TF they only test kernel 1,1 for SAME padding
test_forward.py L381

test_forward.py L385

Other kernels will fail for SAME padding. More info on it. https://discuss.tvm.ai/t/why-we-only-support-kernel-1-1-for-tf-conv2d-transpose-same/4957

[FrozenGene]

Could we support non-1x1 conv_transpose too? I think maybe it is a good time to do it completely no matter tf or tflite.

[apivovarov]

There are two types of padding in TF and TFLite - VALID and SAME.
conv_transpose op does opposite of what conv2d is doing.
conv2d output size is the same or smaller than its input. Bigger the kernel - smaller the output.
conv_transpose is opposite - Bigger the kernel - bigger the output.
We can use any kernel for padding 'VALID' - 2x2, 3x3, etc.
But if we use padding 'SAME' then the output size should be the same as the input.
If kernel is 1x1 - then the output is the same as the input. So, kernel 1x1 is implicitly SAME.
If we increase conv_transpose kernel size then the output will have extra paddings.
In order to make the output size to be the SAME as the input we need to remove padding from the output.
The model which needs conv_transpose op is palm_detection.tflite. It uses VALID padding.
Looks like in most of the cases people use conv2d and conv_transpose with VALID padding.
I think it is fine if we merge the PR with known limitation for padding SAME (it is not that frequently used anyway). We probably should wait till TF frontend supports SAME paddings for non-1x1 kernel and do the same in TFLite frontend.
Maybe they will decide to add VALID/SAME padding field to Relay op directly. In that case we will just pass padding type to Relay op as-is.

[FrozenGene]

Ok.

[FrozenGene]

LGTM

[tqchen]

Thanks @apivovarov @FrozenGene