Shape inference on the Torchvision’s Mask R-CNN causes a segmentation fault

Bug Report

Is the issue related to model conversion?

Probably not.

Describe the bug

When I try to run shape inference on the Torchvision’s Mask R-CNN it causes a segfault. However, check_model doesn’t return any warning/exception.

System information

• OS Platform and Distribution (e.g. Linux Ubuntu 16.04): Linux Ubuntu 18.04.4
• ONNX version (e.g. 1.7): 1.6.0
• Python version: 3.6.10
• PyTorch version: 1.6.0a0+9907a3e
• Torchvision version: 0.7.0a0
• GCC/Compiler version (if compiling from source): N/A
• CMake version: 3.14.0
• Protobuf version: 3.12.2
• Visual Studio version (if applicable): N/A

Reproduction instructions

import torch
import torchvision
import onnx

model = torchvision.models.detection.maskrcnn_resnet50_fpn(pretrained=True)
model.eval()
x = [torch.rand(3, 800, 800)]
torch.onnx.export(model, x, "mask_rcnn.onnx", opset_version = 11)
onnx_model = onnx.load("mask_rcnn.onnx")
onnx.checker.check_model(onnx_model)
onnx.shape_inference.infer_shapes(onnx_model)

Expected behavior

Shape inference doesn’t cause any errors.

[gramalingam]

This is a bit strange. As far as I know, the inference used by onnxruntime is the same regardless of whether you use the python or C/C++ API. There is a slightly different implementation of the inference in onnxruntime and ONNX. But even here, if the behavior is different, then one of them needs to be fixed. This seems worth investigating.

[przybyszewskiw]

It looks like when I use onnx 1.7.0 and onnxruntime 1.4.0, model loads correctly in both APIs. Thanks for your help!