Re: backward compatibility of models saved with 1.2.0
Thanks Thomas for producing a minimal example to reproduce the issue. On Tue, May 29, 2018 at 3:49 PM Thomas DELTEIL wrote: > Created a github issue detailing the problem with a reproducible example: > https://github.com/apache/incubator-mxnet/issues/11091 > > Thanks, > > Thomas > > 2018-05-29 11:50 GMT-07:00 Sergio Fernández : > > > Hi Mario, > > > > I can't give you much details. But it looks there is a bug exporting the > > parameters' names to a JSON models. > > > > I wonder if anybody else in the community has faced this bug. > > > > Cheers, > > > > > > On Mon, May 28, 2018 at 5:42 PM, Marco de Abreu < > > marco.g.ab...@googlemail.com> wrote: > > > > > Hello Sergio, > > > > > > you are right. We are following semantic versioning and thus, every > model > > > produced within the same major version (e.g.1.x) has to be backwards > > > compatible. > > > > > > Could you please provide a small example so we can reproduce this? We > > > definitely do not want our users to retrain their model if they update > > > MXNet. That's a serious issue and we'd love to follow up. > > > > > > Best regards, > > > Marco > > > > > > Sergio Fernández schrieb am Di., 29. Mai 2018, > > 02:35: > > > > > > > Hi, > > > > > > > > I can't find anything related on that in the 1.2.-0-incubating > > changelog, > > > > so I assume models produced by the latest version would be backward > > > > compatible with old versions, such as 1.1.0. But we've found that the > > > > parameter model produced is very different and doesn't load. > > > > > > > > Can you point me to any documentation that could help us to load the > > > model > > > > in 1.1.0 without re-training? > > > > > > > > Thanks. > > > > > > > > > >
Re: backward compatibility of models saved with 1.2.0
Hi Mario, I can't give you much details. But it looks there is a bug exporting the parameters' names to a JSON models. I wonder if anybody else in the community has faced this bug. Cheers, On Mon, May 28, 2018 at 5:42 PM, Marco de Abreu < marco.g.ab...@googlemail.com> wrote: > Hello Sergio, > > you are right. We are following semantic versioning and thus, every model > produced within the same major version (e.g.1.x) has to be backwards > compatible. > > Could you please provide a small example so we can reproduce this? We > definitely do not want our users to retrain their model if they update > MXNet. That's a serious issue and we'd love to follow up. > > Best regards, > Marco > > Sergio Fernández schrieb am Di., 29. Mai 2018, 02:35: > > > Hi, > > > > I can't find anything related on that in the 1.2.-0-incubating changelog, > > so I assume models produced by the latest version would be backward > > compatible with old versions, such as 1.1.0. But we've found that the > > parameter model produced is very different and doesn't load. > > > > Can you point me to any documentation that could help us to load the > model > > in 1.1.0 without re-training? > > > > Thanks. > > >
Re: backward compatibility of models saved with 1.2.0
Hello Sergio, you are right. We are following semantic versioning and thus, every model produced within the same major version (e.g.1.x) has to be backwards compatible. Could you please provide a small example so we can reproduce this? We definitely do not want our users to retrain their model if they update MXNet. That's a serious issue and we'd love to follow up. Best regards, Marco Sergio Fernández schrieb am Di., 29. Mai 2018, 02:35: > Hi, > > I can't find anything related on that in the 1.2.-0-incubating changelog, > so I assume models produced by the latest version would be backward > compatible with old versions, such as 1.1.0. But we've found that the > parameter model produced is very different and doesn't load. > > Can you point me to any documentation that could help us to load the model > in 1.1.0 without re-training? > > Thanks. >
backward compatibility of models saved with 1.2.0
Hi, I can't find anything related on that in the 1.2.-0-incubating changelog, so I assume models produced by the latest version would be backward compatible with old versions, such as 1.1.0. But we've found that the parameter model produced is very different and doesn't load. Can you point me to any documentation that could help us to load the model in 1.1.0 without re-training? Thanks.
