Running ONNX models from Caffe2 doesn't work for any example models

[jeffreyksmithjr]

I attempted to execute the steps shown in the Caffe2 tutorial entitled Run an ONNX model with Caffe2: https://render.githubusercontent.com/view/ipynb?commit=c04c7d721da9b1fa1dd0fe5d5551086c61877586&enc_url=68747470733a2f2f7261772e67697468756275736572636f6e74656e742e636f6d2f6f6e6e782f7475746f7269616c732f633034633764373231646139623166613164643066653564353535313038366336313837373538362f7475746f7269616c732f4f6e6e78436166666532496d706f72742e6970796e62&nwo=onnx%2Ftutorials&path=tutorials%2FOnnxCaffe2Import.ipynb&repository_id=110873948&repository_type=Repository#Run-an-ONNX-model-with-Caffe2

With the squeezenet model it throws the error:

warning: onnx optimizer is unable to parse input model. (The IR version of the ONNX model may be too old.)

So, I also tried every model from the example models repo: https://github.com/onnx/models and none of them worked. Most of them were learned in CNTK and there appears to be a range of issues with running models learned in CNTK from Caffe2.

Caffe2 version is 0.8.dev.2018.04.16. OS is Debian Stretch.

I understand that this is more of a Caffe2 issue to file. Just also filing it here to let you, the ONNX developers know as well.

[jeffreyksmithjr]

Cross-posted: https://github.com/pytorch/pytorch/issues/6644

[bddppq]

@houseroad I think we need to update the squeezenet model file here as well