torch.onnx.export(
                    pose_model,
                    (x,),
                    check_point.split('.')[0] + '.onnx',
                    input_names=["input"],
                    output_names=["output"],
                    opset_version=11,
                    dynamic_axes={
                        "images": {0: "batch"},
                        "keypoints": {0: "batch"}
                    },
                    autograd_inlining=False,
                    operator_export_type=torch.onnx.OperatorExportTypes.ONNX_ATEN_FALLBACK
                )

解决报错：

加上   autograd_inlining=False,就行了