Tests for ConvolutionTranspose3D are probably invalid.

[tsocha]

Describe the bug
Normally test data for convolution is in format:

• [N, C, H, W] or [N, H, W, C] for 2D
• [N, C, D, H, W] or [N, D, H, W, C] for 3D
  but in these tests: https://github.com/Microsoft/onnxruntime/blob/master/onnxruntime/test/providers/cpu/nn/conv_transpose_op_test.cc#L394-L474
  data shape rank is 4 for both 2D and 3D tests.
  3D and 2D tests differ only in rank of strides and amount of pads dimensions which seems invalid.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 16.04.6 LTS
• ONNX Runtime installed from (source or binary): source
• ONNX Runtime version: master
• Python version: n/a
• GCC/Compiler version (if compiling from source): 5.4.0
• CUDA/cuDNN version: n/a
• GPU model and memory: n/a

Expected behavior
Expect test to pass using different data shapes.

[pranavsharma]

Thanks for reporting this. Is your finding consistent with other frameworks? Or the onnx tests?

[tsocha]

Yes I checked onnx tests, pytorch and tensorflow.
In all of these cases they use 5D input for 3d convolutions and pooling.
In the onnxruntime tests I linked above padding might be invalid because there should be 6 zeros(above and below each axis) and in these tests there are 9 zeros.
Gathered data:
ONNX python unit tests input shapes:
convtranspose_3d: [1, 1, 3, 4, 5] and [1, 2, 3, 3, 3]
averagepool_3d: [1, 3, 32, 32, 32]
maxpool_3d: [1, 3, 32, 32, 32]
pytorch-converted.avg_pool_3d: [2, 3, 4, 4, 4]
pytorch-converted.conv_3d: [2, 3, 3, 4, 5]

pytorch conv3d: https://pytorch.org/docs/stable/nn.html#id22
pytorch conv-transpose3d: https://pytorch.org/docs/stable/nn.html#conv-transpose3d
tensorflow conv3d: https://www.tensorflow.org/api_docs/python/tf/nn/conv3d
tensorflow conv3d_transpose: https://www.tensorflow.org/api_docs/python/tf/nn/conv3d_transpose

[hariharans29]

@tsocha - Thanks for reporting this. Yes, that seems right. It doesn't look like a "3D" test case as the test name suggests. Let me add some real "3D" test cases :).

[hariharans29]

Hi @tsocha - I created a PR to rename the misleading test names but I am unable to add 3D tests right away as they are not supported by the implementation yet. Will track that as a separate issue and will close this issue after the PR renaming the misleading test names is merged. Thanks for reporting this.