/hydrus/ - Hydrus Network

Hi hydrus dev!

I did a little bit of work and was able to get some basic neural network tagging to work (with the Illustration2Vec model).

It's very unstable at the moment and freezes up everything with WriteSynchronous, but it is able to assign and commit local tags that are somewhat accurate.

You can see my changes at https://github.com/antonpaquin/hydrus

I don't know if this is even on your radar at the moment so no worries if you don't want to get to it right away, but if you're interested I'm happy to clean up / rework / polish what I wrote.

>>9122

The line I am going with is: I am very thankful to the industry for giving me a year to catch up on my backlog. :^)

It was a weaker than normal year for my gaming interests. Pretty much the only space babes were in Smash or otherwise on switch, and I am neither smashfag nor switchfag (single-player offline pc master race). Most of the characters were too hamfisted or ugly for me to be interested. I still had a great time shitposting with other Anons about it, so I am glad I watched.

Metal Wolf Chaos is great news.

Cyberpunk is obviously potentially breddy gud and a possible exception, and the clear winner of the show, although we haven't seen anything real yet. I don't trust them not to compromise and downgrade for actual release, particularly if MS and Sony push for it to work on current gen (or even next gen) consoles. Still, that trailer showed what their world was about in a clear and exciting and vivacious way that the more mainstream West is unwilling to go for.

Cuphead expansion and Nioh 2 look good, but I still haven't got around to the originals, so that is all more backlog-fodder.

The only other one that really piqued my interest was Satisfactory. I am crossing my fingers it is at least somewhat approaching Factorio-level complexity and hasn't got any always-online shit.

I enjoy business news and politics, so that side was interesting. MS seem to be solving some of their problems, Sony seem to have a little hubris, Nintendo shot their wad.

>>9142

Thanks m8, this looks really interesting. I am impressed you were able to get this up and actually integrated into my shit code. I had a shallow look at your code, but yeah I am not ready to get into this properly until the current downloader work is finished. I never dove too deep into the nuts and bolts of Illustration2Vec–does it use Keras, or did you have to alter the I2V db a bit to make it work on that? Is the tag.py in your .tar.gz your file? How is Keras to work with?

I am a complete fucking sperg about integrating code from other people, but if you were able to figure out a really simple module where I could do simple stuff like:

import nntag

nntag.learnimage( path, learning_data )

tag_info = nntag.predictags( path )

sort of stuff, I'd absolutely add it to hydrus. I don't understand the requirements of the learning loop of ML yet, so I'll need to research that first when I get around to this stuff. I am particularly interested in expanding on the I2V example to create our own, personalised tag dbs, built with our own CPU cycles. If you are read up on how they built their db, I'd be interested in a high-level explanation of what is needed to do this stuff and what data goes back and forth. If you are simply pulling from what they provide, I'd still be interested in the brief version of what happens here–what format and how reasonable are the confidence levels you get with the tags? I see you send a 224x224 image at the model–is that based on what they trained the db with, or is there a magic number in that size?

>>9122

>Be me

>I always like smash

>I'm looking forward to seeing more about it

<40 MINUTES OF SMASH

Just why.

Felt like one of the weakest Nintendo e3 directs in ages. Why wasnt the details about smash pushed to a separate smash direct? Why not release those details through the treehouse with Sakurai sitting in, which they did anyways? Smash's everyone is here trailer did more than enough for hype, a simple couple notes after words that returning mechanics and to play with characters on e3 to notice tweaks would have been an easy hook to get people playing.

I didn't need to see 40 fucking minutes of it going into every little change for every character and no Metroid Prime 4, no retrostudios title, or anything. I would have even taken more indies showcasing something. It really felt like they had practically nothing and banked it all on Smash. I dont even think the e3 BotW direct was this empty.

>>9148

Results:

https://github.com/antonpaquin/Hydrus-Autotagging/tree/master/illust2vec-flask

Classifies with a flask server that drives a parsing script.

Advantage: can work without code changes in current hydrus

Disadvantage: this will only tag one image at a time and takes a few button presses per image

>>9151

>>9148

Thanks for this info. This looks great stuff–if it is all correct with you, I'll point people at your repos here in the v311 release post if they would like to experiment. I am interested in knowing further how these different systems work well and badly.

For the one-image-at-a-time issue, it has long been a thought to convert the file-lookup system to an automated 'figure out any tags for these 10,000 files in idle time please' maintenance routine, particularly once I had introduced the new bandwidth system, which will help us not DDoS the services we are pulling from. Your use case here is another vote in doing something like this, and generally to generalise the way(s) hydrus can pull tag recommendations from other services and scripts.

I'll reiterate that I can't put any time into this atm, but I am interested in having an ongoing conversation about how I can make these workflows better for you. I am keen to get hydrus and The Imageboard Community into the ML game in the coming years.

>>9152

Yes. In making hydrus, I have come to make a point about putting 'human eyes' in front of certain decisions, especially for large automated systems. Figuring out a good workflow is often as difficult as getting the technical side working. It is easy to fuck up a script, and if that script touches 100,000 files, it can be a huge pain to fix. Tag siblings is a good example of this–it seemed simple when I got in, but it turns out for multiple reasons–human preference and complexity of language and translation issues and simple human mistake and technical complexity at the data and gui levels and low CPU availability at certain critical moments in the sibling processing pipeline–to be much more complicated.

If we gain the ability to harvest millions of new tag mappings based on ML systems or otherwise, I would imagine putting them in a separate low-trust cache on the side of the 'current' mappings–probably just in their own service in the hydrus context–until you approved them. The decision to apply these tags to files or not would start being approved by human feedback, with human eyes seeing every decision, until some critical threshold of success were reached, at which point we could trust it.

I am keen to train our own models, so I can imagine us starting with just a handful of new tags, like say 'character:pepe'. The model could inherit its opinion of what is and isn't pepe based on the existing mappings in your db, and then it could ask you about cases in which it was not sure. After refining its model, it can start suggesting pepe for new files. Again, you say yes, yes, no, yes, no, until you are saying yes enough times in a row for it to be 99.8% or whatever confident about pepe, at which point you can say, 'ok, you are ready to pepe anything'. You can then move onto the next tag to train on.

I'll reiterate that I am not read up on this stuff yet, but if it is possible to separate or even quantize these decisions and how they affect the model, I could even write a ML repository where this info could be shared, or even make it exportable to pngs or something, so Anons could share recognition models for different tags and thus share the workload. "I spent ten hours teaching my client to recognise the top ten 2hus so you don't have to." "Here is an official model that determines the difference between 'medium breasts' and 'large breasts'." "Here is a 'feminine penis' detector." And so on, depending on how all this data shakes out, which kinds of tags it works well and badly for, and how people want to use and share it. The public tag repository has been a great success in a bunch of ways, but I think the future has a lot more sharing metadata about what tags are, and less sharing 'this file has this tag'. I am certain these systems will be complicated and require multiple iterations of work to be useful in broad ways.