Mehdi2277

8 ms·

31.
by Mehdi2277 4y ago
My team started with an untyped codebase a year ago. I added a rule to CI that every pr that touches files with type errors must decrease the type errors in touched files by at least 5. When we started we had like 30k type errors. A year la
32.
by Mehdi2277 4y ago
You can write type stubs for other libraries. Over past year my team had a 100k line codebase that we’ve gotten mostly typed. But we still use a lot of other untyped libraries like tensorflow. So in past couple months we’ve started adding t
33.
by Mehdi2277 5y ago
The total memory of the model is less important then the memory needed to compute one batch. I’ve worked with recommendation models used in serving that were 10ish terabytes. The simple trick was most of the memory was embeddings and only s
34.
by Mehdi2277 5y ago
Two towers are not independent of neural networks. Most two tower models I encounter are neural networks for each tower with optionally a mixing towering/layer after the two. And sometimes it's not two but even more towers dependi
35.
by Mehdi2277 5y ago
That's not possible to write in most languages. If a language does support compose operator or custom operators then yes I think that style is good too. But there is no way to write that style in python/ruby. My ideal preference i
36.
by Mehdi2277 5y ago
I agree the approach itself is quite brute force heavy. There is a lot of information that could be used and I’d hope is helpful like grammar or language, traces/simulations of behavior, etc. But when I say alpha code/copilot is g
37.
by Mehdi2277 5y ago
I've also worked in the area and published research in it a couple years ago. I almost worked for a company focused on neural program synthesis but they did a large pivot and gave up a couple years ago working on much simpler problems
38.
by Mehdi2277 5y ago
The difference is your keys are types. Dependent types are about generics where the keys are allowed to be values. You can not in Java directly do, Optional<1> and Optional<“Name”>. You can only have types as arguments. C++ has
39.
by Mehdi2277 5y ago
It can work but generally you either have performance issues or less library support or run into more operational pains. I mainly work on building library tooling over tensorflow. Losing support for good number of operations is not somethin
40.
by Mehdi2277 5y ago
My experience is bulk of the problem is insufficient monitoring. ML systems need heavy monitoring and should be sending lots of metrics to stuff like prometheus/grafana. There should also be validation/consistency checks for all d
41.
by Mehdi2277 5y ago
I currently work as an ML engineer and have interviewed on both sides for some well known companies. The absolute demand in number of people is small compared to popularity. It would not surprise me at all if many computer science master&#x
42.
by Mehdi2277 5y ago
I normally care about reproducible rng results and do try to seed all rngs used. There are lots of applications where randomness is used but you still want repeatable behavior. ML experiments is one situation where randomness is common, but
43.
by Mehdi2277 5y ago
False. Poetry implements it’s own dependency resolver. It does not use pip’s resolver. For most packages this is fine. For a couple packages it’s resolver does not work. Pytorch is one major ml library that is incompatible with poetry as po
44.
by Mehdi2277 5y ago
Good spark support would be a good answer for batch/stream processing. I am a little scared of definition of support. Apache beam supports like 5 runners (flink, spark, data flow, etc) but the quality of runner support is extremely inc
45.
by Mehdi2277 5y ago
Language choices are less about language to me and more about ecosystem of libraries. Python has generally been very strong in the ml/data science realm. I know Julia is catching up but am unsure just how much it covers. Example of pyt
46.
by Mehdi2277 5y ago
I've worked at facebook in past. Facebook has been complimented at a number of interviews and one company I worked for very actively poached fb engineers. It was extremely clear for that company given the resume system even marked resu
47.
by Mehdi2277 5y ago
The systems I worked on were mainly for user curated recommendations and not public rankings. Your rating is only small part of your recommendation. A large amount of your recommendation is finding other users with similar ratings on simila
48.
by Mehdi2277 5y ago
It is trade off. Pure single threaded code is worse in exchange for much better multithreaded code. And the current sentiment in the python dev mailing list looks positive. Previous attempts at Gil removal had much bigger drop in single thr
49.
by Mehdi2277 5y ago
No new race conditions are expected as this work replaces Gil with finer grained performant locks. It still lowers performance of single threaded code some vs no Gil (about 10 percent) but 10 percent drop in single threaded performance is w
50.
by Mehdi2277 5y ago
That’s a workaround that is not as ergonomic as writing python more directly. The person working on this Gil project is one of the major maintainers of a key ml library. They have used c/c++ binding approach but want to make life easie
51.
by Mehdi2277 5y ago
I don’t think counting there vote less is useful. Low quality signal that is plentiful is very often better than sparse higher quality signal. My past experience is working on recommendation systems at tiktok/snap. Having a lot of weak
52.
by Mehdi2277 5y ago
Your first point isn’t true because most of the 5 star usage boils down to 1/5 stars and not 2-4. There are some users that use whole range and the info does have value. The main problem though is while that system may be better for ce
53.
by Mehdi2277 5y ago
There are two issues with 5 star ranking. One the more thought required to give feedback the less likely users give feedback. More interactions is very helpful for ml so minimizing the decisions needed by a user helps a lot. Great ui/u
54.
by Mehdi2277 5y ago
I'm familiar with both. The person who's leading the GIL removal is one of the main authors of one of the two leading ML libraries in python (pytorch) and his motivation is for primarily performance for ML. He explicitly talks abo
55.
by Mehdi2277 5y ago
Yes they do. I’ve written cython/numba as work arounds before. A lot of times if you need a small operation done many times the multiprocessor overhead is bad, but writing a pure python for loop over numpy/other tensors is awful f
56.
by Mehdi2277 5y ago
I use python mostly for numpy/tensorflow/etc. The machine learning ecosystem. That ecosystem cares a lot about multithreaded performance. So historically answer has been write c/c++ and then bind to python. This work is mainl
57.
by Mehdi2277 5y ago
A couple years ago I would have had a bigger interest in this. Pytorch added elastic training as part of version 1.9 when they incorporated torch elastic into it. Tensorflow added an easier way to do fault tolerant training in 2.4 with back
58.
by Mehdi2277 5y ago
I view ANN search for retrieval as clear winner in retrieval methods for domains with very large number of items like O(10 million +) that you want to high amount searches. I've worked at a couple different major tech companies and I&#
59.
by Mehdi2277 5y ago
I’ve worked now at Facebook, tiktok, and snap. None of my teams used tickets so there are no points to track. Most of the time my tasks were just discussed with my team and main way you could track is by prs. How many points a task is worth
60.
by Mehdi2277 5y ago
I’ve worked at a couple big tech companies and one scrum. Notable differences, much less ceremonies, planning is also more up front with deviation allowed but not needing additional planning, and ticketing being optional thing that varies b

More ›