Just wanted to say thanks for the warm welcome from HN when the app was released last month — I hope this blogpost answers the questions that were raised back then.
I’d be happy to answer anything else you’d like to know!
I am making an app that takes pictures and tries to tell you if the food in the picture has allergens. I didn't know if I should feel humble or just laugh. (I decided it was hilarious in the end) But it made me aim higher in a hackathon last weekend. I also use your app in my elevator pitch for people to understand.
I am going to chime in and really suggest they hire a qualified legal professional - if they intend to share this application. A simple disclaimer may not be adequate.
Yea I'm not going to release the app. I was already on the fence on it. I could have finished it months ago but I was worried about legal implications . Thank you .
No problem. I wouldn't even give it away, with any disclaimer, without consulting a qualified legal professional. That sort of thing is just begging for a lawsuit.
I am kind of curious how well you do. If you ever do get it going, there may be a use for it - but, again, liability is a huge factor in something like this.
I am not a lawyer but I have spent a whole lot of time with lawyers and in court rooms. It was part of my business, indirectly. Thankfully, I sold and retired years ago.
Oh, no thank you. I'll just watch from afar. Also, I have a Windows phone. Yup. I know my shame. It's the only Microsoft OS I have. I figure nobody is writing malware for it. ;-)
While it seems great to be able to take a picture, that’s going to be very hard to know what’s in a sauce, stuffed inside of something, complex presentation, etc.
Why not be able to search over a list, and once on an item, show frequently related (e.g. garlic if you search onion). It may not involve any “AI,” but it’ll be far more accurate and easier to implement.
Yes you are right. The app works by identifying the dish and then searching in an ingredient database. If the dish looks visually the same to one with peanuts then the app would have a false negative .
So the long version of this story is that the app was released in partnership with the HBO Go / HBO Now team, and using the terms of service they used for these apps (which are only available in the US & Canada). We’ve been working with lawyers to release the app worldwide without running afoul of any local laws, and I’m keeping my fingers crossed that will get cleared this week, we just gotta make sure our terms of service are up to par… After all, we wouldn’t want to fall prey to something we made fun of this very season ;) http://www.vulture.com/2017/04/silicon-valley-recap-season-4...
Canadian hotdogs on US soil (aka "cross-dogging") and vice versus are disambiguated by shadow inclination vs timestamp to estimate latitude (outdoors) and fingerprinting microfluctuations in the power grid and resulting effects on lighting (indoors, Canada has a much larger share of hydroelectric).
GPS was considered, but rejected as insufficiently cool.
Can you speak to the origins of it being 'not hotdog'?
Did you, or the writing staff or other consultants, start with hotdogs or dick picks?
It made me laugh because it reminded me of this popular lecture(o) that was, and is, passed around tech circles
Any relation? Or fun coincidence?
(o) https://youtu.be/uJnd7GBgxuA?t=5555s .. the lecture is given by andrej karpathy on 2016-09-24 and the timestamp goes to a point in the lecture where the lecturer discusses an interface made to have humans compete with conv nets to identify hotdogs
Ha seems like a fun coincidence. The writers came up with it early in the writing of season 4, and I started working on it sometime in the Summer of 2016 iirc. As far as the origin story goes, it was just great writers coming up with a great joke! Our lead technical consultant Todd mentioned to them we could actually build their joke for real, and the show jumped on the idea!
Not offtopic at all! Dirty hack for sure. The enclosure I bought was a hack, the drivers were a hack, and there was software on top that was a hack as well. But the developer experience was totally awesome… Almost made the constant graphics crashes worth it. Almost.
Probably only 20% of the world's hot dogs are just a basic hot dog with mustard on it. Once you move past one or two condiments, the domain of hot dogs identification along with fixings gets confusing from a computer vision standpoint.
Pinterest's similar images function is able to identify hotdogs with single condiments fairly well:
Having embedded tensorflow for on-site identification is all well and good for immediacy and cost, but if I can't really properly identify whether something is a hotdog vs. a long skinny thing with a mustard squiggle, what good does that do me? What would be the next step up in your mind?
I ask this as someone who is sincerely interested in building low cost, fun, projects.
While we’re here and chatting about this, I should say most of the credit for this app should really go towards the following people:
Mike Judge, Alec Berg, Clay Tarver, and all the awesome writers that actually came up with the concept: Meghan Pleticha (who wrote the episode), Adam Countee, Carrie Kemper, Dan O’Keefe (of Festivus fame), Chris Provenzano (who wrote the amazing “Hooli-con” episode this season), Graham Wagner, Shawn Boxee, Rachele Lynn & Andrew Law…
Todd Silverstein, Jonathan Dotan, Amy Solomon, Jim Klever-Weis and our awesome Transmedia Producer Lisa Schomas for shepherding it through and making it real!
Our kick-ass production designers Dorothy Street & Rich Toyon.
Meaghan, Dana, David, Jay, Jonathan and the entire crew at HBO that worked hard to get the app published (yay! we did it!)
I am glad I am not the only one with questions about the external GPU, I had considered trying that, but came to the conclusion that the data transfer between CPU to GPU would be too slow for ML tasks.
So, what is your opinion on this ? if you had to do it again would you use the eGPU or just use AWS or another GPU cloud service .
My takeaway is that local development has a huge developer experience advantage when you are going through your initial network design / data wrangling phase. You can iterate quickly on labeling images, develop using all your favorite tools/IDEs, and dealing with the lack of official eGPU support is bearable. Efficiency-wise it’s not bad. As far as I could tell the bottleneck ended up being on the GPU, even on a 2016 MacBook Pro with Thunderbolt 2 and tons of data augmentation done on CPU. It’s also a very lengthy phase so it helps that’s it’s a lot cheaper than cloud.
When you get into the final, long training runs, I would say the developer experience advantages start to come down, and not having to deal with the freezes/crashes or other eGPU disadvantages (like keeping your laptop powered on in one place for an 80-hour run) makes moving to the cloud (or a dedicated machine) become very appealing indeed. You will also sometimes be able to parallelize your training in such a way that the cloud will be more time-efficient (if still not quite money-efficient). For Cloud, I had my best experience using Paperspace [0]. I’m very interested to give Google Cloud’s Machine Learning API a try.
If you’re pressed for money, you can’t do better than buying a top of the line GPU once every year or every other year, and putting it in an eGPU enclosure.
If you want the absolute best experience, I’d build a local desktop machine with 2–4 GPUs (so you can do multiple training runs in parallel while you design, or do a faster, parallelized run when you are finalizing).
Cloud does not quite totally make sense to me until the costs come down, unless you are 1) pressed for time and 2) will not be doing more than 1 machine learning training in your lifetime. Building your own local cluster becomes cost-efficient after 2 or 3 AI projects
per year, I’d say.
I have used the AWS machine learning API and would recommend it. The time savings using that vs running it on my hacked together ubuntu-chromebook-mashup is worth more than what I had to pay.
I have also used Paperspace. My only issue was that whatever they use for streaming the virtual desktop to the browser didn't work over sub 4MB/s network connection.
It's interesting how amenable image classification neural networks are to the "take working model, peel off last layer or two, retrain for a new application" approach. I've seen this suggested as working pretty well in a few instances.
I guess the interpretation is that the first few normalize->convolution->pool->dropout layers are basically achieving something broadly analogous to the initial feature extraction steps that used to be the mainstay in this area (PCA/ICA, HOG, SIFT/SURF, etc.), and are reasonably problem-independent.
For sure, although I should say, for this specific instance I ended up training a network from scratch. I did get inspiration from the MobileNets architecture, but I did not keep any of the weights from their ImageNet training. That was shockingly affordable to do even on my very limited setup, and the results were better than what I could do with a retraining (mostly has to do with how finicky small networks can be when it comes to retraining).
Yes, that’s what you see in the picture, although as completely personal advice, I would stop short of recommending it. For one there are arguably better cases out there now, and you can sometimes build your own eGPU rig for less. Finally, the Mac software integration (with any eGPU) is very hacky at the moment despite the community’s best efforts, and I had to deal with a lot of kernel panics and graphics crashes, so overall I’m not sure I would recommend others attempt the same setup.
Nice write up that should become the go-to tutorial for TF and local training. Helped me a lot w/ the mobile part, it was a bit strange to thing about transfer the training when I read at first but it became clear in the second reading.
Pretty fascinating and encouraging to see how much was accomplished with a laptop and consumer GPU. Gave me some great ideas. Also happy to see Chicago dogs properly identified.
Ha you have no idea how hard chicago hotdogs made my life! There was a joke in the show about Dinesh having to stare at a lot of “adult” imagery for days on end to tune his AI, but my waterloo was chicago hotdogs — the stupid pickles end up hiding the sausage more times than not, which makes it hard to differentiate them from say, a sandwich.
Less interesting than you'd expect, as it was for a rapid mobile app prototyping class.
We had a telesync'd demo that let you play along with a Jeopardy episode by yelling answers at your phone. The app knew the timing markers for when the question was asked + when a contestant answered, so would only give you credit if you beat the contestant with the correct answer.
Our model user was "people who yell answers at the screen when Jeopardy is on."
Still think it would have made a decent companion app to the show though...
Trebek's elocution is just something you pick up on after rewatching an episode enough times. He has really interesting ways of emphasizing things, but they seem normal if you're just listening to them once through.
One of my primary motivators behind building this blogpost was to show how exactly one can use TensorFlow to ship a production mobile application. There’s certainly a lot of material out there, but a lot of it is either light on details, or only fit for prototypes/demos. There was quite a bit of work involved in making TensorFlow work well on a variety of devices, and I’m proud we managed to get it down to just 50MB or so of RAM usage (network included), and a very low crash rate. Hopefully things like CoreML on iOS and TensorFlow Lite on Android will make things even easier for developers in the future!
yeah, that's my main pain with the TF docs – great if you just want to try one of the MNIST tutorial variations, but there's a lot more you need to figure out when you get beyond of these "hello world" examples…
This is amazing - impressed by your persistence to source the training data yourself, that must have been tedious!
Did you try quantizing the parameters to shrink the model size some more? If so, how did it affect the results? It also runs slightly faster on mobile from my experience.
Great question — I did not, because I had unfortunately spent all of my data on that last training run, and I did not have a untainted dataset left to measure the impact of quantization on. (Just poor planning on my part really.)
It’s also my understanding at the moment that quantization does not help with inference speed or memory usage, which were my chief concerns. I was comfortable with the binary size (<20MB) that was being shipped and did not feel the need to save a few more MBs there. I was more worried about accuracy, and did not want to ship a quantized version of my network without being able to assess the impact.
Finally, it now seems that quantization may be best applied at training time rather than at shipping time, according to a recent paper by the University of Iowa & Snapchat [0], so I would probably want to bake that earlier into my design phase next time around.
Thanks! Haven't seen that paper, I'll check it out. I think quantization only helps with inference speed if the network is running on CPU with negligible gains on GPU (Tensorflow only supported CPU on mobile last I looked which was a while ago). However your app is already super fast so don't I think anyone would notice if it was marginally faster at this point!
At MongoDB World this past week they did a demo of stitch where they actually built something similar with no back end code required and used the Clarifai API and an angular front end. It took like less than 80 minutes and could like run on prod of I wanted.
Thanks for the kind words! To prevent impostor syndrome, I should clarify that I worked on the app for many, many months — basically since August of last year — as a nights/weekends thing. It’s true that the final version was built almost from scratch in a few weeks, but it wouldn’t have been possible without the time investment in the preceding months. Although for the most part I just wasted a lot of time because I had no idea what I was doing lol (still don’t)
Yes, I was very excited we were able to release it for Android… And even though we used React Native, there were so many native (and C++) bits, it ended up being quite complex!
As for the gear, I think it’s really damaging that so many people think Deep Learning is only for people with large datasets, cloud farms (and PhDs) — as the app proves, you can do a lot with just data you curate by hand, a laptop (and a lowly Master’s degree :p)
Love this architecture. I think Im going to adopt some of it for HungryBot, my nonprofits diet tracking research arm. I think on-phone predictions solves a lot of my affordability issues.
As someone who maintains a popular android camera library; what is this app using to take photos on both iOS and Android? Android can be a bit tricky with device-specific differences and Camera 1 vs. Camera 2 API changes.
The amazing react-native-camera plugin! [0] I’m still getting a few camera-related crashes on Android right now, but overall I would say it makes things pretty smooth!
Well for a while I was lulled into complacency because the retrained networks would indicate 98%+ accuracy, but really that was just an artifact of my 49:1 nothotdog:hotdog image imbalance. When I started weighing proportionately, a lot of networks were measurably lower, although it’s obviously possible to get Inception of Vgg back to a “true” 98% accuracy given enough training time.
That would have beat what I ended up shipping, but the problem of course was the size of those networks. So really, if we’re comparing apples to apples, I’ll say none of the “small”, mobile-friendly neural nets (e.g. SqueezeNet, MobileNet) I tried to retrain did anywhere near as well as my DeepDog network trained from scratch. The training runs were really erratic and never really reached any sort of upper bound asymptotically as they should. I think this has to do with the fact that these very small networks contain data about a lot of ImageNet classes, and it’s very hard to tune what they should retain vs. what they should forget, so picking your learning rate (and possibly adjusting it on the fly) ends up being very critical. It’s like doing neurosurgery on a mouse vs. a human I guess — the brain is much smaller, but the blade says the same size :-/
Not at the moment — although you’ll find the most critical aspects explained in detail in the post. The rest I fear will age very quickly… With stuff like CoreML and TensorFlow Lite on the immediate horizon, I can’t imagine people will want or need to use the cumbersome manual approach I had to use to ship this app. Anything in particular you’d like to see? I can try to share it in a follow-up post or in comments here.
I have a quick question on this. The blog post mentions that you guys went with CONV-BATCHNORM-ACTIVATION (still unsure on whether this is the better order), but in the model code that is posted the batchnorm and activations are the other way around. Which ordering did you end up using?
Thanks! I'm trying out your model on a small data set I've been playing with for identifying invasive species of flowers [0] and so far it's working way better than my initial version that was based on resnet (though slower)!
I’d be happy to answer anything else you’d like to know!
Original thread: https://news.ycombinator.com/item?id=14347211
Demo of the app (in the show): https://www.youtube.com/watch?v=ACmydtFDTGs
App for iOS: https://itunes.apple.com/app/not-hotdog/id1212457521
App for Android (just released yesterday): https://play.google.com/store/apps/details?id=com.seefoodtec...