I'm more excited about StableHLO and IREE than about their integration into Pytorch, Tensorflow, etc.
I want to see a DSL that can be used to describe models elegantly and then export them either to a shared object or to something that can be run with a runtime (in this case IREE). Things like ONNX and TorchScript promised this but I've had little luck getting these to work well enough to trust them in large scale production deployments.
I understand that PyTorch is an awesome tool for researchers, but it doesn't necessarily fit into a prod environment.
> I understand that PyTorch is an awesome tool for researchers, but it doesn't necessarily fit into a prod environment.
You need to write some infrastructure around PyTorch to make it work. Something like a key/mapping in each checkpoint that says which architecture to choose with which parameters.
It sure could be easier, but is saving the model's code into the checkpoint enough? Things like the data pre-processing expected by the model would also have to be included for it to really be self-contained.
Yes I'm facing this when trying to convert a YOLO-based model to TorchScript for mobile (React Native) usage. I wish I can also package the whole pre/post processing from Python to TorchScript instead of having to rewrite it in JS.
I'm curious about your view on ONNX. At work we did a few prototypes and it seemed to work well enough for our use cases, and we're moving to it. What is it that we haven't seen yet that gave you trouble?
Admittedly we're on a reasonably easy situation: we just have to deploy models (some from scikit-learn, some from Keras, some from PyTorch) to various users who mainly run a specific version of python under Windows and Linux, with CPU and GPU support.
Brandon, I’m curious how the goals of your representation differ from, say, Jaxprs.
Why should I look at Aesara’s representation of multi-dimensional array programs when I might already use JAX’s?
Does Aesara support a staging transformation that allows me to construct programs in your representation from a subset of Python?
I’m personally interested in the answers to these questions, given what I know about IREE, JAX, and XLA — as a user in the space, I haven’t been able to determine how Aesara would actually benefit me over JAX.
Note that I know that Aesara can use JAX as a backend — but I’m trying to ascertain what one extra layer buys me.
The problem with this is there do seem to be quite legitimate physicists who disagree with the "sky is falling" conclusion that all popular discussion is premised on. I recently heard an interview with Dr. Richard Lindzen that, for instance, provides what sounds like a very reasonable counter to the prevailing attitude.
Anyone who says "the science is settled, the laws of physics cannot be violated" fails to realize what physics is even about. The vast majority of victories that physics has had since Isaac Newton are based on incredibly simplified models that strip out all of the complexity from a system. Basically everything interesting is treated as linear (as in f(x) ~ x) and a lot of the other stuff is just thrown into constants. After a few decades of working with that, both in theory and with experiments, you might get comfortable enough to add some small corrections to your theories. In this way you step closer and closer to "truth" but we are still very far off, even with our best physics, from fully understanding anything of even moderate complexity. This is not to say that we don't get incredibly useful results from physics (just look at our technology!), but it does mean that we need to constrain statements about how much we really understand and where our "laws" really are applicable. Take for instance the equation for kinetic energy, k=.5 * mv^2. Hugely useful, but dead wrong if you try to apply it to things moving at relativistic speeds.
The problem of climate science, from my perspective, is that we can't strip it down like we do for simple systems like a ball moving in space. Climate is inherently very complex; if you try to ignore how oceans interact with vegetation, how vegetation interacts with clouds, how clouds and rain are connected, etc, then your models can be interesting and might reveal something, but they can't be used to predict what will happen with any degree of confidence because they are too far removed from the real thing. And if you don't strip something out of the system when doing your modeling, well then good luck: you'll never understand anything at all because the thing is too complex to be used to calculate anything of interest.
Sorry, this is typical denial discourse up there. "Oh but we are not perfect, we can't be sure".
The fact that there a greater margin of impredictability should make us even more wary of potential disastrous consequences.
Plus, scientists working on the question do acknowledge that there is a big room of impredicatability. Which is why they make best-case and worst case scenarii that vary wildly.
There's however a consensus on the fact that it's going to be bad. What don't know is how bad. Which is IMHO frightening enough
I'm sorry but you are wrong. I am describing how physics works and it is at odds with many of the claims made regarding our understanding of what is a very, _very_, complex system. You can't just claim there is "consensus that it's going to be bad" without having the sort of scientific discussion I am trying to have. Let's discuss what we know, what the limits are of that knowledge, etc, before we declare that the sky is falling.
I do think you make a great point, which is that the risks and rewards need to be considered. The problem is that analyzing the cost is also a pretty complex problem and there are very real arguments to be made that going green could end up being very costly.
FWIW, I actually came into the climate science debate on your side and have been won over by paying attention to a broader set of sources. One of the things I really can't stand about the climate change disaster is that even if you accept all the claims from the IPCC, it pales in comparison, in my opinion, to other issues that seem a bit simpler to me. The collapsing insect populations, for instance, can have a pretty direct impact on human health and happiness. The climate discussion seems takes all the attention even though no one has made a credible argument as to why it is more dangerous than other problems we face.
edit: I'll add that scientific modelling is _very_ difficult in general, even for much simpler systems than what we are talking about here. These models have tons of simplifying assumptions baked into them and they can be made to spit out just about anything, depending on what values for certain parameters are chosen. The tools aren't particularly trustworthy.
edit: I'll add a bit more. I do think climate change is happening and I do think that humans might have a big part in why it is happening. What I don't find compelling is why this is necessarily an end-of-the-world scenario. The typical argument is: "there has already been warming and there will be further warming. Then there will be runaway effects, then the warming will become extreme and life will become very difficult for most people in most places." Well I do agree with the first part, but it is _really_ hard to predict what happens afterwards, i.e. are there going to be runaway effects. There are reasonable arguments that there won't be. The arguments that there will be tend to focus on a few scientific facts and ignore the complexity of the system at large (i.e. they ignore negative feedback mechanisms and focus only on positive feedback).
Well don't you think that climate change plays a part in a collapsing insect population. I know pesticides are killing them but I have a hard seeing how unpredictable weather patterns, extreme climate events and resulting ecosystem perturbations in their environment may not contribute to that collapse.
The mere fact that an end-of-the-world scenario is possible should mobilize us all. End of the world is not something you gamble on. And, of course, it won't an end of the world, it will a slow and painful life on an unlivable planet.
Can you give me any positive feedback from the release of carbon ?
We've seen this line of reasoning in the COVID crisis. I won't be fooled twice. Everyone is systematically overstating their confidence. I'd rather deal with the problem as it arises than do unnecessary damage.
And I definitely think that relying on the weather a.k.a. "renewable energy" when you have solid prediction that its going to be unreliable, while your prediction of the possibility of averting the crisis is unknown, is beyond dumb. It's suicidal.
We're choosing the worst possible strategy right now. We have the highest probability of the worst case scenario: chaotic climate and weather while relying on said climate for energy generation. Planning for worst case is planning for the crisis happening regardless of your actions, because your confidence about the impact of your actions is pretty low.
There is no metric by which the current suicidal strategy makes sense. We're not minimizing the damage in the worst case. We're not adapting to the situation as it comes. The only scenario in which the current strategy has a better outcome is if we somehow were right about everything, not only about the impact of CO2, but about the geopolitical actions taken by China and India.
The probability that we're on the best possible strategy is practically zero.
Yeah I came here to say this as well. I gave up on LaTeX for my resume something like ten years ago. Since then I use HTML/CSS and render it with wkhtmltopdf.
There were some reasonable bits in here, but there is just enough typical, ridiculous Silicon Valley nonsense to largely turn me off from the article.
> I actually got a lot done
In an article that seems to be promoting the idea of relaxing and taking a break once in a while, the author couldn't resist sneaking in nod to the typical fetishistic obsession for constant productivity that is endemic in this crowd. It seems like we're told that it is possible to relax AND be productive. Is it so unreasonable to disconnect from work every once in a while without worrying about being "productive?"
> Metrics are useful
No they aren't, not for this. There is something seriously wrong with someone if they need a $300 piece of equipment (+ a few hundred more $ for the phone to connect to it) in order to understand something so fundamental about their bodies. We aren't talking a diagnostic imaging to find a suspected tumor here. We're talking about being tired. This is absurd.
I have one of those rings. I bought it because I like gadgets. I think I know what the author is getting at - it's nice to have a soulless device tell you yes, you aren't imagining it, you're having a down day. Take it easy on yourself.
But isn't that the core issue with "these" people: they have lost the ability to be able to interpret their body's sensations and warning signals and burn out as a result?
I'm not excempt from this. But the urge to substitute bodily awareness with sensory gadgets seems to gather religious proportions in some circles.
People can be surprisingly good at blocking out signals, especially when they have learned to regard a signal as broken and unhelpful.
Aron Ralston, the guy that 127 Hours is based on, in his memoir singles out a specific moment in his childhood when he decided to disregard his sense of fear, because it was paralyzing him and stopping him from doing normal things with his friends like skiing. Operating without regard for fear was obviously not a great fix. As an adult he lost friends who refused to go into the backcountry with him because he took irresponsible risks that put them in danger. Then he lost his arm.
People who struggle with depression have the same adversarial relationship with fatigue. When you spend much of your life struggling to force yourself to go when your brain says stop, you start to take a cynical attitude toward your sense of fatigue. It's a chronic liar, a stopped clock, the boy who cried wolf. On the other hand, once in a while it's telling the truth.
Anxiety can present as fatigue as well. When your brain is afraid of something, inducing a sense of fatigue is an easy way to avoid it.
When you can't trust your sense of fatigue, there's no simple rule for handling it. It's dangerous to ignore it, but on the other hand, you can't just trust it, because then you'd spend most of your life in bed. Maybe today the right answer is to "find joy in it" (barf) but tomorrow the right answer is to suck it up and work through it. There's no inspirational slogan that solves it. You just do your best to figure it out every day.
More people need to just work out. If you lift on a regular basis and you're having a day where you don't feel great, you'll know for certain very quickly if it's a "real" down day or not, without needing some special ring, because you'll miss lifts. It's somewhat true of cardio, too, but it's much more possible with extremely low energy output per rep exercise to just power through and do it. With strength training, you'll know for certain when your strength isn't there.
I don't understand. In this article we are imaging widespread and frequent failure of critical infrastructure and we are supposed to further imagine that we're still interested in working on our relatively unimportant software? I suppose there are critical software systems out there, but they're already written so we don't really need to think about what languages they'll be work on with.
No offense intended, but what you describe sounds pretty dismal to me. We have what you describe where I live (France), but I prefer to go to the markets where I know the butcher/green grocer/cheese person/dried fruits and nuts person/fish monger/etc. and they know me. I stand in a line, sometimes for a while, and finally they get to me and ask me how I've been doing recently. Then they ask me for my order and frequently tell me something interesting or useful about what I'm ordering (e.g. where it comes from, how it was grown, etc). They might tell me how to prepare something or offer a suggestion as to what I could buy that I don't usually buy. They might even tell me something about an event happening in the community soon. It's pretty rad.
The human operated lanes still exist in my store. Many people use them, and they do have chats at the checkout line.
And no offense :), but I don't go to the grocery store to make friends. I have much better or more interesting things to do. If I could get fast reliable delivery and not even have to go, I would choose that option.
The point relevant to the OP is that self-checkout systems can be implemented well.
I think the idea of making acquaintance with people in one's community is what has held society together historically, and the derision for human interaction with others (shopping, eating, pub, etc) is the downfall of civil society.
I chat with the cafe owner on my street fairly often.
My grocery store, however, is 20 minutes away in a neighboring town. They have a large staff which is in frequent rotation. The staff also is mostly early 20s or younger age. Add all these things up, and it just doesn't make much sense for me (or them) to go out of our way to socialize.
Yes same here. I really go crazy when technology that takes us further away from each other is presented as some sort of progress. Why don't we look around at how isolated we've become? Is it any wonder that our stupid little factions (right vs left, black vs white, etc) are pulling further and further away from each other when so much effort is being put into limiting our contact with one another?
To be fair, the self checkout are mostly replacing the cashiers in supermarkets that are lowly paid for a very though job, who probably don't know you and who would probably be fired if they "lost" one second saying something nice to you, not the friendly fruit producer from your local market.
Attitudes vary from person to person, but I don't really think being 20 minutes from nature really counts. I lived in Paris until recently and had a really rough time for a few reasons, but the biggest one is was the lack of green places to walk. Whenever I brought it up to anyone, people would say "oh, what about Bois de Boulogne or Bois de Vincennes?" Both were inaccessible by foot from where I lived due to distance and would take 30 minutes to get to by public transit. So what's the idea, then? We should budget an hour in a subway tunnel each day so we can spend some time in a park? That's no way to live; nature isn't something we should treat like a museum. It is something we should feel a part of.
For me, thankfully, there was an easy solution: I moved to one of the towns just outside of the city. Now I have ample access to lots of trees, birds, flowers, etc. If I get on my bike, which I do frequently, I can be in some really beautiful regional parks within 40 minutes or so, and the ride there ain't too bad either.
My favorite place to bike is Parc naturel régional de la Haute Vallée de Chevreuse (French people aren't into the whole brevity thing). I can't speak to hiking but I have to assume there are nice walks to be had in the area because I frequently see people with hiking gear on. You can get there from Paris on the RER B: just take it all the way south to Saint-Rémy-lès-Chevreuse. Île-de-France is, by and large, quite flat, but Chevreuse is a nice valley so you can get some really pretty scenery.
EDIT: I'll add that last fall I was biking around there with my wife and we found some most righteous porcini mushrooms growing on the side of the road. It really is nice out there.
I should also add that nearer Paris is Forêt domaniale de Meudon, which is in the direction of Versailles. Nice forested area. I can't speak to the walking/hiking but I assume it is not bad (I've only been out that way on bike).
