8 ms·
PyTorch Mobile
- Nimitz14 7y agoYes yes yes! I've been anticipating this for so long. However I'm a bit skeptical about doing quantization after training, in my experience you have to do quantization-aware training for there not be a large performance decrease. I guess it works though otherwise they wouldn't have released it?
- tsbinz 7y agohttps://arxiv.org/abs/1906.04721 https://arxiv.org/abs/1906.04721 is an example paper where they perform data-free quantization after training (using batchnorm parameters to get information about the data distribution) without large performance drops.
- jayyhu 7y agoAccording to their docs[1], three different ways of quantizing your model are supported - one of which is quantization-aware training (3). [1] - https://pytorch.org/docs/master/quantization.html#quantization-workflows https://pytorch.org/docs/master/quantization.html#quantizati...
- hhtoyou 7y agoi hate microsoft's way of getting things done! stop reinventing the wheel !
- tmoot 7y agoWell, there goes my weekend :)
- faceshapeapp 7y agoThis is super exciting! I wish they also had browser support without having to go through onnx.
- crubier 7y agoThis is amazing! It might be a personal feeling, but my opinion is that Facebook is SO much better than Google at delivering open source libraries that people want, and support them in the best way. React vs Angular, Pytorch vs Tensorflow. These are just two examples among many where the Facebook framework arrives a bit later on the market but is then supported awesomely and improved continuously while the Google framework arrives earlier but then becomes a hot mess of non retro compatible upgrades and deprecations... My “loyalty” to Facebook open source libraries just keep growing.
- m90 7y agoNot to dismiss anything you mentioned, but looking at the big picture Facebook also published things that went awry / never really gained adoption. Flow, Buck, Hack to name a few. I mean, the cool things are cool (even if I have a really hard time with Facebook as a company) but I think it's a bit of a stretch to say Facebook has "the recipe".
- chrischen 7y agoI would say their track record is still better but it’s hard to compare since some projects are more impactful than others.
- crubier 7y agoI actually have been burnt by Flow so I agree with you, I don’t say that they have “the recipe”. But my subjective point of view is: Google project = “Meh, this big opinionated framework will be probably be abandoned in 2years and all those poor folks who adopted it will have to rewrite everything”, while Facebook project = “Wow they just took the best of the state of the art, and packaged it in a simple reusable library that will be supported decently”. Not a big fan of Facebook as a company, but as an open source contributor they are among the best in my book.
- umanwizard 7y agoIf anyone from the team is reading, can you comment on how much code size this adds to iOS apps?
- shafte 7y agoWatch David's talk coming up (https://www.facebook.com/pytorch/ https://www.facebook.com/pytorch/), it will contain some answers :)
- dzhulgakov 7y agoIn this experimental release with prebuilt binaries it’s about 5Mb per architecture. This includes all operators for inference (that is forward only). We’re working on selective compilation so that you can build a smaller bundle with only a subset of ops that you use. With that for common CNNs it should get to 1-2 Mb range or even smaller.
- bjornjaja 7y agoCurious, what does PyTorch do for embedded when Torch uses Lua? Can’t imagine python for embedded is better than Lua?
- mlevental 7y agoLua torch has been dead for a while https://github.com/torch/torch7 https://github.com/torch/torch7. pytorch doesn't use any Lua.
- reubenmorais 7y agoIt's supposed to go through their JIT first, so there's no Python running on embedded devices. Of course, this means if your model can't easily be made compatible with the JIT, you can't use PyTorch Mobile. But that's not surprising.
- reducesuffering 7y agoDoes this leverage the Neural Engine part of the chip on iOS devices like Core ML does? Can anyone compare using this to Core ML Apple API's?
- sanxiyn 7y agoAs far as I know, there is no way to use Neural Engine without going through Core ML. This does not go through Core ML, hence this can't use Neural Engine hardware acceleration.
- diffset 7y agoWhat advantage does this provide over Apple's CoreML?
- orf 7y agoDoes anyone else have to maintain backend PyTorch based services? Is it just me or is it a complete mess? Members of my team have spent literal months tracking down memory leaks, the performance of these services are always sub-par to Tensorflow based ones and the less said about the atrocious memory/cpu usage the better. What's the advantage of using PyTorch when you have things like Tensorflow Serving ready to productionize any model with ease?
- wil421 7y agoAs a side point, looking at maven and the org.pytorch gave me a cold shudder remembering my java days. How does CocoaPods compare to Maven, NuGet, or NPM?
- Traubenfuchs 7y agoMaven is simple, works extremely well and behaves pretty sane. IDE integration is amazing and unless you write a build/deploy pipeline you don't even need to call a single mvn command. What do you not like about it?
- woolinsilver 7y agoI can't believe he's comparing Maven to npm. I wouldn't piss on npm if it was on fire.
- spicyramen 7y agoThis is a common trend for being second in market, when we see Pytorch and TensorFlow 2.0, TF 2.0 was created to compete directly with Pytorch pythonic implementation (Keras based, Eager execution). Facebook at least on Pytorch has been delivering a quality product. Although for us running production pipelines TF is still ahead in many areas (GPU, TPU implementation, TensorRT, TFX and other pipeline tools) I can see Pytorch catching up on the next couple of years which by my prediction many companies will be running serious and advanced workflows and we may be able to see a winner there.
- mikkelam 7y agoMight be a stupid question, but what's the advantage of using PyTorch Mobile, compared to converting to ONNX-> TF -> TF lite (or something similar)
- sinkpoint 7y agoIt's non-trivial, certain models like GRU has fundamentally different implementation, so they are not cross-compatible. TF also has no incentive to merge any PR back into its code-base for compatibility, and dragged its feet for months. I have spent a long time investigating into this and hit brick walls after brick walls.