yeldarb

27 ms·

271.
by yeldarb 6y ago
Good catch; someone on YouTube called this out this morning and we released an updated dataset (we had also missed the ones attached to the wheels): https://public.roboflow.com/object-detection/mountain-dew-co... Reall
272.
by yeldarb 6y ago
Totally; the model does a good job finding what you teach it to find! On the flip side, someone in the YouTube comments mentioned that we missed the ones on the wheels and license plates. So I added them to the dataset this morning and kick
273.
by yeldarb 6y ago
Is the app available on the App Store?
274.
by yeldarb 6y ago
Cool to see this on the front page! @Dgetman sent me some samples over the summer and we tried them as a team-building activity (we tasted a bunch of stuff before and then after to compare). It was absolutely insane how it turned pure lemon
275.
by yeldarb 6y ago
I had a look at the scoring. It looks like it penalized you pretty heavily for all the "H" letters on the honey. We added a text penalty because it turns out the model can read and so all the top results were people writing the pr
276.
by yeldarb 6y ago
You can click to reveal those if you want. It runs each image through a variety of prompts including screens for racist symbols, nsfw drawings, and nudity. If you send me a link to yours I can let you know what it tripped up. (One of our he
277.
by yeldarb 6y ago
You must have caught us in the middle of the leaderboard update job. If you go back via the homepage it should link directly to your submission for that prompt so you can see your rank.
278.
by yeldarb 6y ago
The neat thing is: there is no training set! The CLIP model[1] takes arbitrary text prompts. So for these we're just sending eg "An illustration of an upside down dinosaur" and it returns a similarity value between the image&
279.
by yeldarb 6y ago
Update 2: appears this was actually a bug in my code that applies a text penalty :blush: These blocks of color weren't scoring highly for the prompt... they were scoring really low for similarity to text... which ended up giving them a
280.
by yeldarb 6y ago
Nice! We're recalculating all the leaderboards so no worries; check back in a couple hours to see your new ranking.
281.
by yeldarb 6y ago
Update: "An illustration of an empty canvas filled with color" seems to work pretty well to extract the solid blocks of colors from the actual drawings. We're adding that to our prompts now and then will recalculate the leade
282.
by yeldarb 6y ago
Hey everyone! We made this as a hackathon project over the weekend to experiment with OpenAI’s new CLIP model. https://openai.com/blog/clip/ Glad you’re enjoying playing with it and discovering failure modes! Yest
283.
by yeldarb 6y ago
The cool thing about this model is that it actually doesn’t need to be trained! The way to tweak it is just to frame your question in the right way. We’re doing to update the prompts probably to penalize it for similarity to “a solid block
284.
by yeldarb 6y ago
Complexity is a cool thing to ask it about and adjust the score based on. I’m concerned it’d start loving scribbles though. Any other ideas for ways we could help it understand what we’re asking of it?
285.
by yeldarb 6y ago
This is an interesting observation. The cool part about CLIP is that we can ask it any prompt we want. So “fixing” this loophole is just a matter of prompt engineering. We already added a penalty to the score for text by subtracting similar
286.
by yeldarb 6y ago
*Tools used* Another goal of this weekend was to try some products from a few friends of ours. Apart from OpenAI's CLIP model, Firebase (Auth, Storage, Hosting, Firestore), Docker, and AWS Lambda we used the following tools: * Parade[6
287.
by yeldarb 6y ago
*Neat things we learned* We posted the minimal working version on twitter yesterday afternoon and discovered a few things that led to v2 late last night: * Surprisingly, CLIP can read! It was doing a good job of ranking the initial drawings
288.
by yeldarb 6y ago
We're really excited about CLIP[1], OpenAI's new zero-shot image classification model. We think it's going to power a ton of creative new products. But being inspired isn't as fun as being inspiring so this weekend our t
289.
Show HN: Paint.wtf, a drawing game judged by OpenAI's zero-shot image classifier(paint.wtf)
26 points by yeldarb 6y ago|6 comments
290.
by yeldarb 6y ago
This breakdown of the internals may be more what you were hoping for: https://blog.roboflow.com/a-thorough-breakdown-of-yolov4/
291.
by yeldarb 6y ago
To complicate matters further, the code for this improvement of YOLOv4 (published by one of the original YOLOv4 authors) is actually a fork of the YOLOv5 repo: https://github.com/WongKinYiu/ScaledYOLOv4/tree/y
292.
by yeldarb 6y ago
Is this a reference to a specific Startup School video?
293.
How to try CLIP: OpenAI's new universal zero-shot image classifier(blog.roboflow.com)
8 points by yeldarb 6y ago|0 comments
294.
by yeldarb 6y ago
The best thing I've read on positioning is April Dunford's "Obviously Awesome"; highly recommend. I read it after seeing her speak at a conference. I think this might be the same talk: https://www.youtube.com&
295.
by yeldarb 6y ago
> What matters more practically is a comparison between an M1 Macbook Pro and a leading competing laptop with a discrete nvidia GPU. 100% agree. We would have liked to do that comparison but, as mentioned in the article, the required sof
296.
by yeldarb 6y ago
(Disclaimer: I was involved in the prep for this article so I'm probably biased but) I think the comparison between the two 13" MBPs is the most like-for-like. They're similarly priced and outfitted. I'd expect the futur
297.
by yeldarb 6y ago
No, but would love to see it.
298.
Individuals or Teams: Who’s the Better Customer for SaaS Products?(sacks.substack.com)
1 points by yeldarb 6y ago|0 comments
299.
by yeldarb 6y ago
As long as possible. Our jobs usually finish well within the limit but the top 1% hit the limit. One example we've been wrestling with is a merge operation. Usually it's merging about 1000 records which completes in a few seconds.
300.
by yeldarb 6y ago
Anyone know why there's a hard limit of 15 minutes for Lambda (and 9 minutes for Google Cloud Functions)? Still seems really weird to me.

More ›