Onnxruntime fail to run Fasterrcnn_resnet50_fpn test "Cannot split using values in 'split' attribute"

[ghost]

🐛 Bug

I am trying to run the new test script " test_onnx.py " added to vision/test folder (bellow is google colab example with all details to reproduce the same error) . More specifically, I am trying to export Fasterrcnn_resnet50_fpn as ONNX but it is not working.

To Reproduce

Code Steps to reproduce the behavior:

• I am sharing Google Colab notebook .. it generates the same issue: https://gist.github.com/moured/a18c07f847e63eb4811710e4d702ecdf

Error message :

---

Fail Traceback (most recent call last) in ----> 1 a.test_faster_rcnn()

/media/mmrg/vision/test/test_onnx.py in test_faster_rcnn(self) 342 output_names=["outputs"], 343 dynamic_axes={"images_tensors": [0, 1, 2, 3], "outputs": [0, 1, 2, 3]}, --> 344 tolerate_small_mismatch=True) 345 346 # Verify that paste_mask_in_image beahves the same in tracing.

/media/mmrg/vision/test/test_onnx.py in run_model(self, model, inputs_list, tolerate_small_mismatch, do_constant_folding, dynamic_axes, output_names, input_names) 49 if isinstance(test_ouputs, torch.Tensor): 50 test_ouputs = (test_ouputs,) ---> 51 self.ort_validate(onnx_io, test_inputs, test_ouputs, tolerate_small_mismatch) 52 53 def ort_validate(self, onnx_io, inputs, outputs, tolerate_small_mismatch=False):

/media/mmrg/vision/test/test_onnx.py in ort_validate(self, onnx_io, inputs, outputs, tolerate_small_mismatch) 68 # compute onnxruntime output prediction 69 ort_inputs = dict((ort_session.get_inputs()[i].name, inpt) for i, inpt in enumerate(inputs)) ---> 70 ort_outs = ort_session.run(None, ort_inputs) 71 for i in range(0, len(outputs)): 72 try:

/media/mmrg/company-env3.7/lib/python3.7/site-packages/onnxruntime/capi/session.py in run(self, output_names, input_feed, run_options) 140 output_names = [output.name for output in self._outputs_meta] 141 try: --> 142 return self._sess.run(output_names, input_feed, run_options) 143 except C.EPFail as err: 144 if self._enable_fallback:

Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Split node. Name:'' Status Message: Cannot split using values in 'split' attribute. Axis=1 Input shape={1,14328} NumOutputs=5 Num entries in 'split' (must equal number of outputs) was 5 Sum of sizes in 'split' (must equal size of selected axis) was 17910

Expected behavior

ONNX runtime successfully does inference.

Environment

• Latest OnnxRuntime
• Latest Torchvision
• Latest Pytorch
• Ubuntu 18.04 - 64x
• Python 3.7 & I also tried it on 3.6 same error
• CPU / GPU

[hariharans29]

It seems there has been an issue with the conversion to ONNX. 14328 cannot be split into 5 equal parts and hence the complaint by the runtime. Can you pls check with the model conversion project and raise a bug there ?

[stale[bot]]

This issue has been automatically marked as stale due to inactivity and will be closed in 7 days if no further activity occurs. If further support is needed, please provide an update and/or more details.

[pustar]

ersion project and raise a bug there ?

met the same problem, have you solved it? how? thanks

[twoapples1]

Fasterrcnn_resnet50_fpn cannot achieve dynamic reasoning in the batch dimension, but it can achieve dynamic reasoning in the h and w dimensions. If you want to achieve multi-batch reasoning, you need to fix the batch size when export the model and then perform static reasoning in the batch dimension.I hope this can help you