ShapeOptimizer and TransposeOptimizer produces wrong graphs.

[xadupre]

Describe the bug

Attached file: a non-optimized graphs where TransposeOptimizer (_slice_handler) and ReshapeOptimizer.

debug_noopt.zip

TransposeOptimizer

It fails at that particular node Transpose:

onnxruntime can load the non optimized graph (equal to numpy.fft.fft2(numpy.random.rand(3, 1, 4).astype(numpy.float32), (1, 4)). After optimization, it returns the following error:

FAIL : Node (Ma_MatMul__358) Op (MatMul) [ShapeInferenceError] Incompatible dimensions for matrix multiplication

ReshapeOptimizer

It produces the following error stack:

  File "tensorflow-onnx\tests\test_backend.py", line 179, in _run_test_case
    return self.run_test_case(func, feed_dict, [], output_names_with_port, **kwargs)
  File "tensorflow-onnx\tests\backend_test_base.py", line 403, in run_test_case
    self.assert_shapes_correct(g, self.config.allow_missing_shapes, not self.config.skip_onnx_checker)
  File "tensorflow-onnx\tests\backend_test_base.py", line 326, in assert_shapes_correct
    self.assertTrue(tf2onnx_shape is not None)
  File "lib\unittest\case.py", line 682, in assertTrue
    raise self.failureException(msg)

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):
• Tensorflow Version: 2.5
• Python version: 3.9
• Opset 13

[TomWildenhain-Microsoft]

@xadupre did you double check that this error occurs using the tools/onnx-optimize.py script? I tried and could not reproduce the error. The shape info might differ if you got these originally from a rfft unit test.

[xadupre]

I got it from a unit test. I tried with the optimizer tool and it fails too.

  File "tensorflow-onnx\tf2onnx\optimizer\optimizer_base.py", line 41, in optimize
    graph = self._optimize(graph)
  File "tensorflow-onnx\tf2onnx\optimizer\transpose_optimizer.py", line 157, in _optimize
    return self._apply_optimization(graph, self._optimize_at_current_graph_level)
  File "tensorflow-onnx\tf2onnx\optimizer\optimizer_base.py", line 62, in _apply_optimization
    graph = optimize_func(graph)
  File "tensorflow-onnx\tf2onnx\optimizer\transpose_optimizer.py", line 170, in _optimize_at_current_graph_level
    if self._handle_nhwc_tranpose(n):
  File "tensorflow-onnx\tf2onnx\optimizer\transpose_optimizer.py", line 324, in _handle_nhwc_tranpose
    return op_handler(trans, p)
  File "tensorflow-onnx\tf2onnx\optimizer\transpose_optimizer.py", line 725, in _unsqueeze_handler
    mid_shape = [shape[p] for p in new_perm_inv]
  File "tensorflow-onnx\tf2onnx\optimizer\transpose_optimizer.py", line 725, in <listcomp>
    mid_shape = [shape[p] for p in new_perm_inv]
TypeError: 'NoneType' object is not subscriptable

[xadupre]

I fixed the bug in PR #1649. In two cases, the shape of a node was empty and the code did not expect that.