Hi Andrew, Your issue corresponds with the current development path of HTM.java - and so while you might not be aware, I have been working on your issue in a more wholistic manner - approaching it from a standpoint of enabling serialization in general. The holidays slowed us down a bit, but I wanted to update you on what's been happening so far with this.
Currently, I'm submitting a PR [1] which handles an issue which is key to having a predictable stable environment with which to test the serialized state of the algorithms; which was the indeterminacy found in successive executions of the TemporalMemory. That is handled with the above mentioned PR now. Also I've been looking into serialization schemes (API libraries such as Cap'N Proto, and Kryo) for the Network - though I think initially I will get plain old java serialization working first to gain some experience as to how to approach more optimized serialization via specialized libraries later (on an architectural level). Starting today, I will now see about getting the java serialization up and running and submit a PR for it. Then later this week I'll be looking into and starting development on a more optimized approach. Happy Holidays! David [1] https://github.com/numenta/htm.java/pull/373 On Mon, Dec 28, 2015 at 10:08 AM, Andrew Dillon <[email protected]> wrote: > Hey David, > > I attached a couple of classes that I am using in my program to manage > Networks and serialize/deserialize Connections. The code is a bit rough, > but I ran through and put in some comments to explain some things. These > classes are used by a GUI program to allow a user to interact with the > Network, but I didn't include any of the UI related stuff. Also, I've used > a few different kinds of data as input to the network (feeding it in using > the feed() method in my HTMNetwork class), but I've primarily used Sine > wave values for simplicity while working on the serialization and > deserialization aspects of the program. > > Andrew > > > On 12/24/2015 2:09 PM, cogmission (David Ray) wrote: > > No problem Andrew, I knew what you meant... ;-) Thanks! > > On Thu, Dec 24, 2015 at 2:07 PM, Andrew Dillon <[email protected]> > wrote: > >> I just realized I said I tagged them with the Serialized interface - I >> don't know if that interface exists - I meant the to say the Serializable >> interface. It's a minor point, but I don't want to cause any confusion. >> >> Thanks again, David, and Happy Holidays to you as well! >> On Dec 24, 2015 1:41 PM, "cogmission (David Ray)" < >> [email protected]> wrote: >> >>> Hi Andrew, >>> >>> Thank you for sending over the files. I will take a look and get back to >>> you in the next couple of days. >>> >>> Happy Holidays! >>> >>> David >>> >>> On Thu, Dec 24, 2015 at 12:46 PM, Andrew Dillon < >>> <[email protected]>[email protected]> wrote: >>> >>>> I attached the classes I tagged with the Serialized interface, but I'll >>>> list them below as well. >>>> >>>> - *Connections* >>>> - *Cell* >>>> - *Column* >>>> - *DistantDendrite* >>>> - *Pool* >>>> - *ProximalDendrite* >>>> - *Segment* >>>> - *Synapse* >>>> - *FlatMatrixSupport* >>>> - *SparseBinaryMatrixSupport* >>>> - *SparseMatrixSupport* >>>> - *SparseObjectMatrix* >>>> >>>> The reason I found it necessary to tag the other eleven classes besides >>>> *Connections >>>> *was to include all of the *Connections*'s members in the >>>> serialization process. Some of the classes I tagged are not directly used >>>> as members of *Connections*, but I needed to serialize them because >>>> they were superclasses of certain *Connections* members, and the only >>>> way to deserialize the subclass without serializing the superclass would >>>> have been to add a no-args constructor to the superclass. This would result >>>> in missing data upon deserialization, though, due to certain fields in the >>>> superclass not being initialized. This quick jguru article >>>> <http://www.jguru.com/faq/view.jsp?EID=34802> might explain what I'm >>>> saying better. >>>> >>>> Thanks for the assistance and no worries about a late reply. I wasn't >>>> even expecting a response today, since it is Christmas Eve. >>>> >>>> On 12/24/2015 11:19 AM, cogmission (David Ray) wrote: >>>> >>>> Hi Andrew, >>>> >>>> Welcome aboard! :-) >>>> >>>> First let me say thanks for using HTM.java, it's very nice to hear >>>> about user experiences indeed. Perhaps it would be easier if you attached >>>> the classes that you altered so that we can mock up our own example to see >>>> what's going on - such as the Connections.java file and whatever else was >>>> necessary to alter? (I would be surprised to find that you had to alter >>>> anything else, actually?). >>>> >>>> As this is the next thing on HTM.java's agenda, this is very >>>> interesting indeed... Also, as it is x-mas eve, I may not be able to get >>>> back to you as promptly as I otherwise would - but please send over the >>>> files as soon as you are able because I am anxious to play with them! ;-) >>>> >>>> Cheers, >>>> David >>>> >>>> On Thu, Dec 24, 2015 at 10:51 AM, Andrew Dillon < >>>> <[email protected]>[email protected]> wrote: >>>> >>>>> Hello all. I've been learning about HTM for a couple of months now, >>>>> reading On Intelligence and Numenta's papers, as well as watching their >>>>> videos. I just began actually working with htm.java (I'm not very familiar >>>>> with Python, so I am unable to use that version) a week or two ago, so I'm >>>>> no expert on it, but I have been able to create some working >>>>> demonstrations >>>>> of it. >>>>> >>>>> I have, however, run into a bit of a problem with saving networks. I >>>>> am aware that htm.java does not currently support this type of operation ( >>>>> <https://github.com/numenta/htm.java/wiki/Call-To-Arms> >>>>> https://github.com/numenta/htm.java/wiki/Call-To-Arms), so I am >>>>> attempting to develop a basic method of saving and recreating my networks >>>>> myself. What I have done thus far is modify a couple of classes to >>>>> implement Java's Serializable interface, in order to save my Layer's (I am >>>>> just working with one right now) Connections object. I have succeeded in >>>>> serializing and deserializing the Connections object, and putting it back >>>>> into a new Layer with the Layer.using() method. >>>>> >>>>> The problem is that when I feed the network (that is using the >>>>> deserialized Connections) the same data it had learned to recognize before >>>>> I serialized it, it no longer predicts the proper values. Its output looks >>>>> exactly like a new network; as though my saved Connections is being >>>>> overwritten or ignored somehow. I've spent the past few days trying to >>>>> figure out what is happening, digging around the source code and trying a >>>>> few different things, but have been unable to produce any results. Do any >>>>> of you folks have any idea how I might go about resolving this issue? >>>>> >>>>> I am sure code samples would be of interest here, but I'm not sure >>>>> what, specifically, I should include as my program is of a decent size. If >>>>> anybody would like some samples, please mention what general >>>>> functions/areas of my program you would like to see, and I'll be happy to >>>>> oblige. >>>>> >>>>> Thanks in advance for any help. I am very fascinated by this project >>>>> and HTM theory in general. I really appreciate what you all are doing and >>>>> that this project was made open source! >>>>> >>>>> >>>> >>>> >>>> -- >>>> *With kind regards,* >>>> >>>> David Ray >>>> Java Solutions Architect >>>> >>>> *Cortical.io <http://cortical.io/>* >>>> Sponsor of: HTM.java <https://github.com/numenta/htm.java> >>>> >>>> [email protected] >>>> http://cortical.io >>>> >>>> >>>> >>> >>> >>> -- >>> *With kind regards,* >>> >>> David Ray >>> Java Solutions Architect >>> >>> *Cortical.io <http://cortical.io/>* >>> Sponsor of: HTM.java <https://github.com/numenta/htm.java> >>> >>> <[email protected]>[email protected] >>> <http://cortical.io/>http://cortical.io >>> >> > > > -- > *With kind regards,* > > David Ray > Java Solutions Architect > > *Cortical.io <http://cortical.io/>* > Sponsor of: HTM.java <https://github.com/numenta/htm.java> > > <[email protected]>[email protected] > <http://cortical.io/>http://cortical.io > > > -- *With kind regards,* David Ray Java Solutions Architect *Cortical.io <http://cortical.io/>* Sponsor of: HTM.java <https://github.com/numenta/htm.java> [email protected] http://cortical.io
