Models are not serializable

[haifengl]

Models are not serializable

[haifengl]

I use XStream (http://xstream.codehaus.org) for serialization of models. Can you have a try? Thanks!

[haifengl]

Previously, all models support Serializable interface. However, I remove them due to several concerns. For example, the exact format is hard to keep stable, class changes can easily make your serialized data unreadable, reading/writing the data in non-Java code is almost impossible. I am still looking for a simple and flexible solution. Currently, XStream seems good enough and your classes don't need implement any interfaces. In the future, I would like to support PMML for some models.

I am very open to this question and would like to listen all kinds of suggestions! Thanks!

[nopper]

Fair enough. Thanks for the suggestion ;) Really appreciated. I am gonna close the issue.

[haifengl]

Can you please reopen this ticket? I would like to listen to different voice. Hope more people will see this ticket. Thanks!

[haifengl]

Thanks!

[kid1412z]

Protostuff(https://code.google.com/p/protostuff/) or Hessian(http://hessian.caucho.com/doc/hessian-serialization.html)? I use them to compress the loaded Objects in my LSH of SimHash version. It reduce the memory of old generation of jvm. In my case, the serialized objects of Protostuff are smaller.

[haifengl]

Thanks! I will check out them.

[haifengl]

Protostuff looks very good. But seems we need create schema for every model. A lof of work.

[kid1412z]

I think schema is used to keep compatible if the model has to be changed.
Otherwise, runtime schema can be used and it doesn't need too much code as shown in the example code.

[haifengl]

Sounds great! Will learn more about it. If we choose it, would you like to work together on it? Thanks!

[kid1412z]

Of course :)
2015-03-25 21:52 GMT+08:00 Haifeng Li [email protected]:

Sounds great! Will learn more about it. If we choose it, would you like to
work together on it? Thanks!

—
Reply to this email directly or view it on GitHub
#20 (comment).

[haifengl]

With runtime schema, it seems that the usage of ProtoStuff is similar with XStream. I prefer not to modify the model object for specific serialization library. So it is better for the users to choose whatever they want. If we finally choose some serialization library, I still want to do it in a separate module and leave smile-core unchanged.

[haifengl]

A second thought: do we really need backward/forward compatibility? If the model member variables change, it usually means a serious change to the algorithm. It is better to retrain the model rather than importing the old model.

[haifengl]

Any comments (or complains) to my second thoughts? Thanks!