MXNet 1.5.0 is not downward compatible to MXNet 1.4.1


I created a Github issue 26 days ago about the failure of loading trained models which still remains unanswered:

I tested it with version mxnet-cu100-1.5.0b20190718 today, and the problem still persists.

Apparently I’m not the only one who experienced problems when loading trained models in MXNet 1.5.0.

Assuring downward compatibility between direct version updates is an important quality feature for any library, especially in the case of Deep Learning. People might have spent several days or weeks training certain models and might be unable to update their version because of this.


The issue has been resolved. :slightly_smiling_face:

Apparently additional safety checks have been added when loading a model in version 1.5.0 and were previously ignored.