[04:09:13] I’ve started exploring captions, descriptions, depicts, and categories to see how things work. I’ll share my observations soon… If anyone is interested in contributing, I’d be happy to collaborate. (re @super_nabla: Seems great! Maybe you could try embedding the caption/description of the subset of images appearing in at least a wiki.) [11:55:11] gopavasanth: yes, that has been already tried. We (VGG - a research group at University of Oxford) have a demo search engine to ~1/2 of all images in wikimedia commons (this demo is not public because we found a lot of NSFW content). We also have a public demo that only includes wikimedia media of the day at https://meru.robots.ox.ac.uk/motd/ [11:58:59] gopavansanth: beyond semantic search on the visual content, we also have search for faces and objects through face embeddings, and audio content alone [12:00:40] (with the telegram bridge I have no idea if I'm sending messages to someone still in the room) [12:01:33] (Your messages are coming through to Telegram.) (re @wmtelegram_bot: (with the telegram bridge I have no idea if I'm sending messages to someone still in the room)) [12:06:11] Jan_ainali: I meant that because on the IRC side I only see a wm-bb user, I have no clue if gopavasanth, the person who I was replying, is still on the telegram room [12:08:27] Ah. Telegram groups don't work in the same way as IRC rooms though. Even if they closed the app, the message will still be there when they come back. [12:11:10] ah! That's good to know. And I guess I also need to @ username instead of username at start of message [12:11:21] @gopavasanth see my reply above ^ [13:06:15] Sounds great. I think for offline preprocessing, you need a machine equipped with some GPU. On a CPU-only environment you hit soon a scalability limitation. (re @gopavasanth: I’ve started exploring captions, descriptions, depicts, and categories to see how things work. I’ll share my observations soon… ...) [14:09:04] super_nabla: unless you are only doing this to tag images only (and I think it would be a waste to stop there) a GPU for online processing of the query would also be useful [14:19:16] Yes, of course it would! But I'm not sure WMF offers this capabilities to volunteers. ToolForge is certainly CPU-only. (re @wmtelegram_bot: super_nabla: unless you are only doing this to tag images only (and I think it would be a waste to stop there) a GP...) [14:31:23] super_nabla: I could wear I read there were AMD GPU available (at least I already started asking around for AMD GPU to test with ROCm because of this) [14:34:25] super_nable: here it is https://wikitech.wikimedia.org/wiki/Machine_Learning/AMD_GPU I ended up in this page while reading about toolforge on wikitech but I guess nowhere there it says that those GPUs were available on toolforge [14:57:21] I object 🤔 (re @jeremy_b: any objections to disabling all emoji reacts for everyone? or are there more fine grained options I can't find?) [18:46:17] That's super interesting.. Thanks for sharing this @carandraug! Much helpful. Will explore further and get back on it! (re @wmtelegram_bot: gopavasanth: yes, that has been already tried. We (VGG - a research group at University of Oxford) have a demo sear...) [18:48:34] I wonder if this spammer understood the irony of their react to your objection. (re @IsmaelOlea: I object 🤔) [18:50:31] Yea you do... I'm around on telegram.. thanks for sending though.. (re @wmtelegram_bot: (with the telegram bridge I have no idea if I'm sending messages to someone still in the room)) [19:00:41] @carandraug Is the software behind https://meru.robots.ox.ac.uk/motd/ open source? If so, could you please share the source code link? [19:57:51] gopavasanth: yes. All our work is open source. The code is at https://gitlab.com/vgg/wise/wise/ with the specifics scripts to create that demo at scripts/wikimedia/motd/ . The step that is most intensive is the extract-features.py (several days depending on your GPU) but if you're interested we could maybe find a way to share that computed data? Or maybe start with a smaller subset of the data [20:05:23] Thanks for sharing the codebase! I’d definitely be interested in reusing this work if sharing the computed data is possible. (re @wmtelegram_bot: gopavasanth: yes. All our work is open source. The code is at https://gitlab.com/vgg/wise/wise/ with the specifics ...)