Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
    Sam Hodge
    @samhodge
    I think I may know the solution to my problem
    I need to save parts of the model and put the pieces together in C++
    Anirudh Subramanian
    @anirudh2290
    i know that for backend symbolic api one uses mx.sym.variable and then binds it to the model
    Sam Hodge
    @samhodge
    that way I can have a resolution independent model
    Anirudh Subramanian
    @anirudh2290
    i have to see how to do that in gluon
    Sam Hodge
    @samhodge
    I can just reshape the appropriate symbole
    Anirudh Subramanian
    @anirudh2290
    yes, but you need something equivalent to mx.sym.variable in gluon correct ?
    Sam Hodge
    @samhodge
    I think so
    it feels like I am trying to rough out an idea at this stage
    Sam Hodge
    @samhodge
    Actually dont worry about that, I think I need to continue on with a fixed resolution model first and simply try that out with the feedforward in C++ and then build it up from there, I am a little confused about if it is possible to build the gram matrix and the inspiration network in C++, but as a start it might be worth just loading up the model and params that I have for a decent trained model.
    I just worry that I am delaying a decision
    Anirudh Subramanian
    @anirudh2290
    so the link that you pointed to https://github.com/samhodge/incubator-mxnet/blob/master/cpp-package/include/mxnet-cpp/symbol.hpp is just the cpp package, its just another frontend like python.
    Sam Hodge
    @samhodge
    Yup, but C++ is easier to deploy than Python
    Anirudh Subramanian
    @anirudh2290
    i am just wondering why you can not keep weight and height as symbol variables and bind it at runtime.
    width*
    Sam Hodge
    @samhodge
    Sounds great
    but I am not sure if I understand how to do that
    how can you load on the params if the width and height are not known
    so when you do x = mx.sym.var('data')
    y = style_model(x)
    Sam Hodge
    @samhodge
    can you also do some sort of magic like a= mx.sym.var('width'), b=mx.sym.var('height') y=style_model(x,height=a,width=b)
    is that what you mean?
    Anirudh Subramanian
    @anirudh2290
    yes
    Sam Hodge
    @samhodge
    I am honestly a bit of a noob
    Anirudh Subramanian
    @anirudh2290
    i am also a noob with gluon
    Sam Hodge
    @samhodge
    its all good you learn by making mistakes
    let me try this out.
    Anirudh Subramanian
    @anirudh2290
    but with symbolic api you just use placeholders for data, and then bind it to the data at the end, i think this should be possible with gluon too
    Anirudh Subramanian
    @anirudh2290
    apache/incubator-mxnet#6087 you probably need something like this
    Sam Hodge
    @samhodge
    I am hoping I can work out how to get the width and height as symbols to the network
    Sam Hodge
    @samhodge
    I think this relates to my issue
    Sam Hodge
    @samhodge
    @
    Anirudh Subramanian
    @anirudh2290
    @samhodge this should help you
    apache/incubator-mxnet#9893
    Sam Hodge
    @samhodge
    thanks I need to go a few steps back at the moment.
    it seems that the training the model that I am trying to serialise as a symbolic network doesnt work without any modifications
    I will put in a ticket now about this issue
    Sam Hodge
    @samhodge
    apache/incubator-mxnet#9989
    So I would be happy with a fixed resolution for now, but I cannot even get that working.
    Sam Hodge
    @samhodge
    @zhanghang1989 do you have an opinion?
    Lutz Roeder
    @lutzroeder

    Screen Shot 2018-03-07 at 8.23.49 PM.png

    Netron now supports MXNet -symbol.json models. Feedback is welcome.

    Anirudh Subramanian
    @anirudh2290
    @lutzroeder awesome thanks a lot ! i am not sure there are enough people from the community here. i will post this in the slack channel. i know that some people have asked for this in the community.
    ThomasDelteil
    @ThomasDelteil
    Thanks @lutzroeder, really cool, I'll use netron in my next talk to show the model architecture ! Just tried it out with the Crepe model, I am just wondering what makes a convolution show as dilates=(1,) rather than just hiding the dilation factor? (I am assuming dilates=(1,) is the same as no dilation?)
    Lutz Roeder
    @lutzroeder
    @ThomasDelteil The default for dilate is (1,1). Any insights if (1,) is some special encoding are welcome. Currently the app shows the values present in the file. It filters defaults for other formats but they have to be added to the operator file as the app doesn’t depend on the MXNet runtime directly. Feel free to open an issues and will have a look.
    ThomasDelteil
    @ThomasDelteil
    I see, thanks @lutzroeder
    Lutz Roeder
    @lutzroeder
    @ThomasDelteil Added a few heuristics for some common defaults and pushed an update.