Questions regarding documentation

Hello, I have some questions concerning the development of DL4J.

I have been following the project for 3 months now and it is very difficult to get involved in the project.

Often outdated information is given. Be it an outdated e-mail address in the README of the repository or links in the navigation bar of the website.

To the website:
The website is outdated and so are the guides, which should make it easier to get started. Instead, you have to leaf through issues, comb through the grid to see if anyone has had a solution.

Another useful piece of information would be to indicate which version a piece of documentation refers to and the date when it was last updated. As can be deduced from issues, “latest” may not always have been the last version.

There are also links to a book, although I would recommend to give a hint that probably most of it doesn’t work as described anymore because the API has changed.

Fortunately, there is an example repository with several examples, which is really very useful. For a beginner, however, the examples are sometimes too complicated or require looking up the documentation, which unfortunately is not reliable and only leads to despair and anger because it doesn’t work as it is stated there.

The release notes are also often incomplete. For example, you suddenly can’t enter “true” as a parameter for backpropagation. It does not appear anywhere. Or that the UI server is no longer scale based is mentioned and that you need to change the Maven dependency accordingly. But that you need an additional dependency “deeplearning4j-vertx” is again only detectable by searching around and finding an issue or pull request.

I also see that the project is being worked on. It seems to me that only one person is working on the project and is more than overloaded with work (see feedback time and number of open issues). An official roadmap would also be appropriate.

I apologize if I have become subjective here. Therefore my real question: When can I expect updated guides? Without them, the framework is very tedious and troublesome for people who have never worked with it before.

Thanks for your feedback. Lots of things you’ve mentioned are kind of known issues for us and we’re trying to improve.

Can you please give us more points to what you find disappointed? Especially related to examples and guides?

As for number of people working - no, it’s far from one person. It’s just typically 1 person that merges pull request from our development repository to primary eclipse repository. That’s where most of work happens: https://github.com/KonduitAI/deeplearning4j

As for guides update: you mean deeplearning4j.org website i suppose?

Hello,

Thank you for the very quick response.

Unfortunately I did not write down the points, but I will do it soon if I find another deficiency.

Yes, I mean the website https://deeplearning4j.org/.

A simple example what had driven me crazy: https://deeplearning4j.org/docs/latest/deeplearning4j-quickstart

Here .backprop(true) is applied. But the parameter no longer exists. For someone who wants to understand the code completely, the parameter cannot be ignored. Maybe it is important. However, since it is not possible to see which version the guide refers to, you may want to consult the Release Notes, which describe that the call is outdated. Often the release notes are not described sufficiently and then the search begins in issues, PRs, grids etc. That’s why it would be a really big help to add the used version in the guide or documentation etc. And instead of latest it would be nice to also see which version refers to latest.

I see… Unfortunately at this moment we don’t control the website, so it wasn’t updated since September i think :frowning:

Oh, ok. So will this site be updated in the future? It sounds odd that you are not in the control of the site. :confused:

At least we definitely hope that it will be :slight_smile:

Sorry to say that I have the same feeling about this project I’ve been following since June 2019.

  • No clear or expressed roadmap.
  • Weekly answers to technical questions in forum (thanks Alex). Lot of recurrent questions about the same problems that never gets clearly explained. Are you really active ?
  • Outdated website and examples with no clear way to give feedback for improvement. For example beta 4 is still the reference version on (https://deeplearning4j.org/). This even generates questions in forum, just for an outdated label!

I tested DL4J for inference and I’ve been successful on several models. So inference is ok for me.
I tested DL4J for training and I’m still wondering if it is ok. At least I was not as successful as for inference in training my model. I asked for help with no answer yet. The performance were not as good as Python/Keras (a factor 2 for my tests, but that can certainly be tuned a bit).

I hope you can release an 1.0.0 version, but I still do not see a strong community of users.

Hello! I would like to give a brief update on the status of things.
Many of you might not know me, but I am the founder of dl4j and started the project back in 2013. I stepped back behind the scenes over time to focus on other parts of the business we run, but am keeping a closer eye on things again. I apologize, as we, like any other OSS project, have had our ups and downs over the years.

I would first of all like to apologize for the status of things in the last few months. A lot has happened on our side if it wasn’t obvious we have migrated away from gitter. I do want you to know that dl4j is alive and well.

Dl4j’s current development flow switched to a dual development/release repo with the development repo being:


and the release repo being:

Regarding your concern about recurring questions, our migration from gitter is exactly why we are doing that. Discoverability of answers was always a problem and one we needed to fix. Support is scattered across stack over flow using the dl4j tag and github issues. Our hope is that discourse will allow for a better long term user experience.

Regarding the community, we have and do have quite a few users in product, but we have not done a lot on the community development side (mostly due to lack of resources till recently). We need to do a better job (and are hiring for this!) to do community development and get more people to talk about their work with dl4j.

We also have recently joined the board of the eclipse foundation: https://www.eclipse.org/org/foundation/directors.php
and are becoming much more active in community development.

The project’s resources have also switched to predominately working on a more multi language future with most of our logic moving down to c++: https://github.com/KonduitAI/deeplearning4j/tree/master/libnd4j

We will be pushing our “tensorflow” useful for model import and other applications with our framework “samediff”.

Regarding:

I asked for help with no answer yet. The performance were not as good as Python/Keras (a factor 2 for my tests, but that can certainly be tuned a bit).

Please, whenever you write a micro benchmark, actually scrutinize if you did it right. Writing benchmarks is harder than people realize. Good advice can be found on the web like here:

There is also our pages on performance:
https://deeplearning4j.org/docs/latest/deeplearning4j-config-performance-debugging
https://deeplearning4j.org/docs/latest/deeplearning4j-benchmark

Please report issues you see ASAP on github issues if you don’t feel like you are being heard, we tend to be responsive there. We have pro active monitoring now for when people post on this forum, stack overflow, or github issues now. Please let us know if there is anything we can do to help.

Shoot me a message if you have any other concerns. Thank you for your feedback and we hope to turn things around for the better this year!

4 Likes