[Frontend][PaddlePaddle] Add topk op and Fix bug, when the output is a dimension, it …

[woobinw]

There are some operation in PaddlePaddle that output a tensor which shape is [1], meanwhile, the slice op have decrease_axis attr, but it should not be squeezed.

@heliqi
@jiangjiajun
cc @junrushao

[tvm-bot]

Thanks for contributing to TVM! Please refer to the contributing guidelines https://tvm.apache.org/docs/contribute/ for useful information and tips. Please request code reviews from Reviewers by @-ing them in a comment.

• No users to tag found in teams: frontend, paddlepaddle _{See #10317 for details}

_{Generated by tvm-bot}

[heliqi]

LGTM

[heliqi]

You can add a few different cases, example:

1. k is a tensor and not just a int
2. set largest/axis/sort attribute
3. The output is only the values or the indices.

[woobinw]

according to paddle.topk api, there seems is no a argument to control the output that only have the values or indices

[heliqi]

Yes, the paddle.topk api will definitely have two outputs.

[jiangjiajun]

LGTM

[junrushao]

Would you like to fix the CI so we could get it in?

[woobinw]

Would you like to fix the CI so we could get it in?
E File "/workspace/tests/python/frontend/paddlepaddle/test_forward.py", line 1356, in slice5 (* user code *) [2023-01-05T11:12:37.644Z] E b, c, h, w = paddle.shape(inputs) # add decrease_axis
I think this error may be related to the version of the paddlepaddle

[heliqi]

Would you like to fix the CI so we could get it in?
E File "/workspace/tests/python/frontend/paddlepaddle/test_forward.py", line 1356, in slice5 (* user code *) [2023-01-05T11:12:37.644Z] E b, c, h, w = paddle.shape(inputs) # add decrease_axis
I think this error may be related to the version of the paddlepaddle
@woobinw
The paddlepaddle version we matched before was 2.1.3.
If you want to upgrade and match the latest paddlepaddle version 2.4, you can pull request and modify the Paddle version in the code and the code.

[woobinw]

Would you like to fix the CI so we could get it in?
E File "/workspace/tests/python/frontend/paddlepaddle/test_forward.py", line 1356, in slice5 (* user code *) [2023-01-05T11:12:37.644Z] E b, c, h, w = paddle.shape(inputs) # add decrease_axis
I think this error may be related to the version of the paddlepaddle
@woobinw
The paddlepaddle version we matched before was 2.1.3.
If you want to upgrade and match the latest paddlepaddle version 2.4, you can pull request and modify the Paddle version in the code and the code.

@heliqi I modified the Paddles version configuration of the above two location. but the following prompt information of code is mismatch 2.4.1 version of Paddle. Can you take a look at this question for verify_model(topk2, input_data=input_data) ?

[2023-01-09T07:38:22.462Z] tests/python/frontend/paddlepaddle/test_forward.py:81: in verify_model [2023-01-09T07:38:22.462Z] baseline_outputs = baseline_model(*[input[:] for input in input_data]) [2023-01-09T07:38:22.462Z] /venv/apache-tvm-py3.7/lib/python3.7/site-packages/paddle/fluid/dygraph/layers.py:902: in __call__ [2023-01-09T07:38:22.462Z] outputs = self.forward(*inputs, **kwargs) [2023-01-09T07:38:22.462Z] /venv/apache-tvm-py3.7/lib/python3.7/site-packages/paddle/fluid/dygraph/io.py:1191: in __i_m_p_l__ [2023-01-09T07:38:22.462Z] return _run_dygraph(self, input, program_holder) [2023-01-09T07:38:22.462Z] /venv/apache-tvm-py3.7/lib/python3.7/site-packages/paddle/fluid/dygraph/io.py:754: in _run_dygraph

[woobinw]

@junrushao
The ethosu test seems to have nothing to do with PaddlePaddle frontend.

[junrushao]

Thank you for pinging me @woobinw! Indeed the failure is not relevant to the change you brought in, and I'm not sure if it's flaky or consistently reproducible error. Retriggered the test just now, and let's see if it could work this time

[junrushao]

@tvm-bot rerun

[junrushao]

Could you rebase to latest HEAD? I’m sure it’s going to pass this time

[woobinw]

Could you rebase to latest HEAD? I’m sure it’s going to pass this time

Yes, It has been rebased to latest HEAD.

[junrushao]

seems like there is some issues in rebasing and now the PR has inflated to 25k lines

[junrushao]

Thanks for the contribution! It’s merged!