Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

YOLOv5: State-of-the-art object detection at 140 FPS

blog.roboflow.ai

Post reply on HN
Print view View original

131–136 of 136 posts

Re: YOLOv5: State-of-the-art object detection at 140 FPS

#131

Post by bcatanzaro »

joshvm wrote: post #34
bcatanzaro wrote: post #5

Why benchmark using 32-bit FP on a V100? That means it’s not using tensor cores, which is a shame since they were built for this purpose. There’s no reason not to benchmark using FP16 here.

Not sure about the benchmark, but the code includes the option for mixed precision training via Apex/AMP.

If you click around enough you’ll see they benchmarked in 32-bit FP. Glad they have a mixed precision training option but I really think it’s a mistake in 2020 to do work related to efficient inference using 32-but FP.

The problem is that your conclusions aren’t independent of this choice. A different network might be far better in terms of accuracy/speed tradeoffs when evaluated at a lower precision. But there is no reason to use 32-but precision for inference, so this is just a big mistake.

Re: YOLOv5: State-of-the-art object detection at 140 FPS

#132

Post by sillysaurusx »

sillysaurusx wrote: post #26
sillysaurusx wrote: post #18

We made a site that lets you collaboratively tag a bunch of images, called tagpls.com. For example, users decided to re-tag imagenet for fun: https://twitter.com/theshawwn/status/1262535747975868418 And the tags ended up being hilarious: https://pbs.twimg.com/media/EYXRzDAUwAMjXIG?format=jpg&name=... (I'm particularly fond of https://i.imgur.com/ZMz2yUc.png ) The data is freely available via API: https://www.tagpls.c…

Well, that didn't take long – our API endpoint keeled over. Luckily, you can fetch all the data directly from firebase: # fetch raw tag data $ curl -fsSL https://experiments-573d7.firebaseio.com/results/.json > tags.json $ du -hs tags.json 14M tags.json # fetch tag metadata (colors, remapping label names, possibly other stuff in the future) $ curl -fsSL https://experiments-573d7.firebaseio.com/user_meta/.json > tags_…

Heh, $1 per day? Try $1k per day. https://imgur.com/duugqHK

Re: YOLOv5: State-of-the-art object detection at 140 FPS

#133

Post by sillysaurusx »

nerderloo wrote: post #109
sillysaurusx wrote: post #18

We made a site that lets you collaboratively tag a bunch of images, called tagpls.com. For example, users decided to re-tag imagenet for fun: https://twitter.com/theshawwn/status/1262535747975868418 And the tags ended up being hilarious: https://pbs.twimg.com/media/EYXRzDAUwAMjXIG?format=jpg&name=... (I'm particularly fond of https://i.imgur.com/ZMz2yUc.png ) The data is freely available via API: https://www.tagpls.c…

How does this have anything to do with the linked article?

The idea with the site is that you can tag your own datasets, and then get the data suitable for yolo training. We've done that ourselves to train an anime hand detector, and other users have reported similar successes. I could've been a bit clearer about that.

Re: YOLOv5: State-of-the-art object detection at 140 FPS

#134

Post by sillysaurusx »

sillysaurusx wrote: post #132
sillysaurusx wrote: post #26

Earlier quoted context omitted.

Well, that didn't take long – our API endpoint keeled over. Luckily, you can fetch all the data directly from firebase: # fetch raw tag data $ curl -fsSL https://experiments-573d7.firebaseio.com/results/.json > tags.json $ du -hs tags.json 14M tags.json # fetch tag metadata (colors, remapping label names, possibly other stuff in the future) $ curl -fsSL https://experiments-573d7.firebaseio.com/user_meta/.json > tags_…

Heh, $1 per day? Try $1k per day. https://imgur.com/duugqHK

We've confirmed that this was someone running `while true; curl ...`, resulting in a $3,700 bill. https://twitter.com/theshawwn/status/1271365062913961984

I guess "be gentle" means "please troll us."

Re: YOLOv5: State-of-the-art object detection at 140 FPS

#135

Post by sillysaurusx »

dang wrote: post #65
sillysaurusx wrote: post #18

We made a site that lets you collaboratively tag a bunch of images, called tagpls.com. For example, users decided to re-tag imagenet for fun: https://twitter.com/theshawwn/status/1262535747975868418 And the tags ended up being hilarious: https://pbs.twimg.com/media/EYXRzDAUwAMjXIG?format=jpg&name=... (I'm particularly fond of https://i.imgur.com/ZMz2yUc.png ) The data is freely available via API: https://www.tagpls.c…

You should post this as a Show HN!

Okay! Thank you. We appreciate the encouragement.

It looks like an HN user on an EC2 server decided to fetch data from our firebase as quickly as possible, running up a $3,700 bill. Once (or if) that's sorted out, and once we verify tagpls can handle HN's load without charging thousands of dollars, we'll add an "about" page to tagpls and submit it.

Post reply on HN