Where communities thrive


  • Join over 1.5M+ people
  • Join over 100K+ communities
  • Free without limits
  • Create your own community
People
Repo info
Activity
    Shivangi Motwani
    @shivangi952_twitter
    Hi! Ya I was able to run the tests successfully with python3. Thanks!
    krinsman
    @krinsman
    Dear @JeanKossaifi : would it be possible to add a section to the installation instructions (in the documentation) explaining how to install tensorly with conda? I figured it out, although it was somewhat of a headache because conda skeleton pypi tensorly didn't work using version 3.7.2 of conda-build (I'm not sure if this is a bug with conda-build or not; anyway it was possible to work around). Here is a github repository I created containing the tensorly package I created and explaining how to install it: https://github.com/krinsman/tensorly_with_conda That being said, it would probably be easier for users if there were an official tensorly conda package, e.g. in the conda-forge channel https://conda-forge.org/#add_recipe or in a private channel of one of the developers, since I'm not confident that the recipe I created is the best possible.
    Also I was unsure whether to raise this as an issue on Github, since some projects have "issues" with the documentation, or if I should just mention it here. If you would prefer this be raised as an issue on Github, please let me know.
    Jean Kossaifi
    @JeanKossaifi
    Hi! That's actually something i've been meaning to do for a while without actually finding time.. I do encourage users to use conda so it makes sense to have a recipe for it.
    In short that will be a great addition!
    Happy to chat here about the details and if you'd like we can work together on a Pull Request to add the recipe + doc (since you've already done work on this)
    I'll create a conda-forge for tensorly
    krinsman
    @krinsman
    I would be glad to work on a Pull Request with you! However I'm not entirely sure how to do that. My concern also in making a Pull Request is that the README I wrote was overly detailed (I think), since it went into specifics of version-control/virtual-environments which aren't necessary for just installing the package. Likewise the recipe I wrote is likewise overly restrictive in controlling the versions of the dependencies (which probably shouldn't be necessary, since there are so few dependencies, and they are all well-maintained projects not likely to often introduce breaking code).
    My point being that the documentation and/or recipe I wrote seem so far off from the final goal of what you would want in the project's documentation that I'm not sure it would be helpful to make a Pull Request based on that (although I understand vaguely that pull requests can be heavily edited before any merges into master are ever made).
    krinsman
    @krinsman
    Also my guess would be that the final thing one would like is to have a conda-forge package, then to just add a line in the installation instructions portion of the documentation that says: "Run conda install -c conda-forge tensorly to install with conda." Since the actual conda-forge package hasn't been made yet though that probably wouldn't make sense to add to the documentation any time soon (although I guess I could do that for you now, and then you could edit then possibly even merge the request into master after you've created the package).
    UPDATE: I just made the pull request, so of course feel free to edit it as much as you want, it is after all your project. After you create a conda-forge for tensorly, it seems we agree such an addition to the documentation might be helpful to conda users.
    Jean Kossaifi
    @JeanKossaifi
    Thanks!
    I packaged it for the anaconda repo, so amended your commit slightly and merged!
    Hopefully I packaged it correctly and there won't be any issue
    I'll push the conda config file soon
    krinsman
    @krinsman
    That sounds great, thank you! Hopefully other people besides me will appreciate it too!
    Jean Kossaifi
    @JeanKossaifi
    let me know if you spot any issue with the package or the conf. file! :)
    krinsman
    @krinsman
    I haven't had the time to test it, so my concern is probably unfounded, but will it work to leave out mxnet as one of the required dependencies in the recipe?
    Jean Kossaifi
    @JeanKossaifi
    I didn’t add it as to not make it a hard dependency (users might want to use NumPy or pytorch instead) and because it would be mean creating several versions due to cpu and GPU support (mxnet-cu8, cu9, etc)
    krinsman
    @krinsman
    You're right (of course). I thought MXNet was the default backend, not numpy, based on what it said in this notebook: https://github.com/JeanKossaifi/tensorly-notebooks/blob/master/01_tensor_basics/tensor_manipulation.ipynb
    If that is a mistake in the notebook then it was the only mistake I found. Anyway, having downloaded, installed, and ran the package from the tensorly channel, I can confirm that it does work.
    By the way I really like the notebook -- it's like Kolda and Bader's survey paper but interactive and with dogs.
    Jean Kossaifi
    @JeanKossaifi
    You're right, the notebook needs to be updated! I wrote it while MXNet was the default but it turned out to be best for users to keep NumPy as default.
    Thanks :)
    I'm planning to write more but it's pretty time consuming!
    Feel free to contribute bits / suggest ideas/changes! :)
    SamJohannes
    @SamJohannes
    Hi @JeanKossaifi and all, I'm planning on implementing a version of randomised ALS for CP decomposition (https://arxiv.org/abs/1701.06600) for a project I'm working on. Would there be any interest in adding this to the tensorly package?
    Jean Kossaifi
    @JeanKossaifi
    Absolutely!
    Happy to look at the code, too, if you have it somewhere on Github!
    SamJohannes
    @SamJohannes
    Excellent! I haven't started yet, but I've got a fork on Github that I'll put my code in. When I've got something decent I'll let you know
    SamJohannes
    @SamJohannes
    I have a question regarding the normalize_factors() method in the the decomposition module. I can't seem to find it being used in any other methods, and it isn't defined in the API reference http://tensorly.org/stable/modules/api.html#module-tensorly.decomposition. Is the plan to eventually include this in a version of parafac(), or is it a standalone function?
    SamJohannes
    @SamJohannes
    Which I guess leads into my next question, there is mention in parafac() of an option to return weights, but this hasn't been implemented yet. I see @JeanKossaifi and @cswiercz are refactoring this file, are you in the process of implementing this?
    Jean Kossaifi
    @JeanKossaifi
    Hey @SamJohannes, sorry for the delay. It is a standalone function. To keep the return arguments the same, parafac doesn't normalise the weights: you can get the normalised factors by calling normalise_factors on these.
    SamJohannes
    @SamJohannes
    No worries, do we want to make this function publicly accessible?
    SamJohannes
    @SamJohannes
    I've made a PR with my work on the randomised ALS algorithm. If anyone wants to look at it, that would be great!
    Jean Kossaifi
    @JeanKossaifi
    Awesome @SamJohannes, I'll have a look!
    Jean Kossaifi
    @JeanKossaifi
    Did you implement CPRAND or CPRAND-MIX?
    SamJohannes
    @SamJohannes
    Just CPRAND so far
    Jean Kossaifi
    @JeanKossaifi
    There seems to be conflicts with master that need to be manually resolved
    SamJohannes
    @SamJohannes
    Really? I've merged master, and github is telling me there's no conflict with the base branch. Am I missing something?
    Jean Kossaifi
    @JeanKossaifi
    It says this on the PR page:
    This branch cannot be rebased due to conflicts
    Rebasing the commits of this branch on top of the base branch cannot be performed automatically due to conflicts encountered while reapplying the individual commits from the head branch.
    SamJohannes
    @SamJohannes
    Am I looking at the correct page? I can't seem to see it on tensorly/tensorly#53
    Jean Kossaifi
    @JeanKossaifi
    yes that's the one :)
    Zongyi Li
    @wumming
    Hello Jean, may name is Zongyi Li. I am the intern student of Anima. Maybe I can contribute!
    Here is the functional/ spectral tensor train method https://arxiv.org/abs/1405.5713
    Jean Kossaifi
    @JeanKossaifi
    Hi Zongyi
    Would be great to have you as a contributor, this is also a function I wanted to add! :)
    Akshay Kulkarni
    @akshaykvnit
    tensorly/tensorly#104 - According to this issue on GitHub, there is a feature request for implementation of parafac2. If no one is working on this, I could contribute to this. Kindly let me know.
    Jaydeep Borkar
    @jaydeepborkar
    Hi Jean, I'm interested in learning about tensors and contributing to TensorLy. I'm about to start learning from these Jupyter notebooks (https://github.com/JeanKossaifi/tensorly-notebooks), followed by the user guide and examples. Do you recommend me anything else as a beginner to get a good understanding of tensors? I watched the talk on Role of Tensors in Machine Learning by Prof. Anima at GTC 2019 to get started. Also, do you recommend me reading the TensorLy paper before jumping to the notebooks? Thanks :)
    Jean Kossaifi
    @JeanKossaifi
    @akshaykvnit sorry about the long delay -- I completely missed this! Would be great if you are still interested in working on this, happy to help along the way!
    @jaydeepborkar The paper is a good place if you want some quick background and references, and it's a short read. The notebooks are a great place to start experimenting, which is probably the best way to learn!
    Jaydeep Borkar
    @jaydeepborkar
    @JeanKossaifi Thank you so much! I'm following the paper right now and wish to keep you updated about my progress. Looking forward to contributing my best :)