TL;DR

OpenCV is fast, but torchvision is faster.

https://arxiv.org/abs/2605.08731

But there is still a huge room for improvement in terms of performance, as for some low level operations StringZilla or Numkong are faster, for some, especially for float32 images, numpy is the best.

The most annoying component is that OpenCV is limited to input shapes like (H, W, C), which limits its application to videos and volumes with shapes (X, H, W, C)

Comments URL: https://news.ycombinator.com/item?id=48421858

Points: 865

# Comments: 148

Submitted 3 weeks ago: https://github.com/pytorch-fdn/ecosystem/issues/67

I work on one of the projects in the list, need to update a link to the project, as old one is not actual anymore. And unclear how to do it => at least with respect to my project Albumentations, the landscape is outdated :(

--- Also, added the project to the Pytorch Ecosystem many years back, but if you ask me about practical value of being the part of the Ecosystem, I would not be able to tell you anything useful.

Systematic, efficient.

Played this game myself. And I did it when moved to the US with a limited English and lack of understanding of the local culture and traditions.

After a few years of dedicated practice, moved me from the state that author describes to the complete lack of fear talking to strangers, I can easily make nearly ever conversation warmer, deeper and more relaxed.

------

A couple more comments, based on personal experience:

[1] It works better if place where you meet is your deep comfort zone, a very familiar place

- gym, if you are going there for some time, know where each type of equipment is. - dance venue that you were going dancing for a while - art class - etc

[2] It helps a lot if you are quite proficient in the activity, expertise brings respect, and higher social status by itself, even when you do not talk to anyone.

[a] in the gym ideal technique > strength > looks / size of your muscles.

- Third class in powerlifting, based on Soviet grading system is a threshold, passing which life changes (question of months, maybe a year). You get more respect from men and curiosity from women, and you get more confident, because you got stronger: https://www.sportscategory.info/en/powerlifting - As your shoulders get broader, fat fat percentage goes down - it improves your appearance -> your confidence -> helps as well.

[b] Dance venue is a great place to meet people and address your fears / issues. Rule of the game - during the class before the social part teacher makes you switch partners => you will be forced to introduce yourself to the partner, this person cannot turn away and will need to reply, introduce themselves.

Later when social part starts - people switch partners every dance => - you start with inviting for a dance people whom you already met during the introductory class. - In 3 hours of social dancing you dance with 20+ people - As your skill grows (question of weeks-months) and dancing with you is not torture anymore, but quite the opposite - it is enjoyable => you get more relaxed, people want to dance with you => conversations start all the time - In dancing, as a man you lead, and this transfers to other activities (helped to become a lecturer teacher in University), but you also better lead the conversation. I.e. it is not a random exchange of information anymore, but you can vary it's direction and emotional component.

--- [3] Some places are better than others.

It is good to go to the gym, to get more friends, but not directly. I do not like talking to people in the gym, I suspect that other people as well.

you are recovering between sets, focussing on the audiobook, moving weights - you are always busy with something. I also heard that women do not like talking to men in the gym as they may feel "no in the best form", i.e. for her - talking to men feels comfortable, when she took shower, picked a cloths that fits her, not when she is sweaty, struggling with weights and sees other ladies in the gym who are more fit.

Places like:

- climbing gym <- very social activity where you solve same problems - trying to climb a route. You can just tell someone who struggled to climb a bouldering problem something like: "Nice!", "Good job!", "Well done", and ask for a tip.

Ot if you already climbed it - give a tip yourself. These are natural openers.

If you climb similar level of problem, you will get stack in the gym in the same spots, taking a break between tries - universe will force you to talk and socialize.

- Dance venue, as I mentioned above - Hikes - any types of group classes: scuba diving, wine tasting, art classes, etc will do the job quite well

Despite augmentation being used everywhere, most discussions are still very surface-level (“flip, rotate, color jitter”).

In this article I tried to go deeper and explain:

• The *two regimes of augmentation*: – in-distribution augmentation (simulate real variation) – out-of-distribution augmentation (regularization)

• Why *unrealistic augmentations can actually improve generalization*

• How augmentation relates to the *manifold hypothesis*

• When and why *Test-Time Augmentation (TTA)* helps

• Common *failure modes* (label corruption, over-augmentation)

• How to design a *baseline augmentation policy that actually works*

The guide is long but very practical — it includes concrete pipelines, examples, and debugging strategies.

Would love feedback from people working on real CV systems.

Link: https://medium.com/data-science-collective/what-is-image-aug...

The motivation was practical rather than theoretical:

Which of these geometric structures, if any, actually survive discretization, noise, and SGD-style training in modern machine learning?

In physics, global and coordinate-free formulations were not aesthetic choices; they were forced when local reasoning stopped working. A recurring structural pattern was:

structure -> symmetry -> invariance -> dynamics -> observables

In modern ML we increasingly see analogous issues:

* parameter symmetries and large quotient spaces * non-Euclidean data (graphs, meshes, manifolds) * highly structured hypothesis classes * training dynamics that are not well-described by flat Euclidean optimization

Some geometric ideas clearly paid off (e.g. equivariance via group actions). Many others did not. I’m trying to understand where future leverage might still lie, and where geometry collapses to interpretation or preconditioning.

Below is my current (incomplete) map of where modern geometry already shows up in ML, or plausibly could.

1. Geometry of data (base spaces)

Manifolds, stratified spaces, graphs and meshes; discrete differential operators (graph Laplacians, discrete Hodge theory); topological summaries (persistent homology).

This seems strongest for representation, spectral methods, and diagnostics. The open question is how much of this geometry can couple dynamically to training, rather than remain preprocessing or analysis.

2. Geometry of hypothesis spaces (architectures)

So far the most successful direction:

* symmetry and equivariance via group actions * quotienting hypothesis spaces * convolution as representation theory * SE(3)- / gauge-equivariant models * architectures encoding invariants or conservation laws

Here geometry restricts the hypothesis class before optimization. I suspect there is still room beyond global groups, toward local gauge structure, fiber bundle–valued representations, and architectures defined by connections rather than coordinates.

3. Geometry of parameters and optimization

Optimization on manifolds (Stiefel, Grassmann, SPD cones), structured or low-rank parameterizations, information geometry and natural gradients.

This seems most effective when constraints are hard and geometric. In looser settings, much of this reduces to preconditioning. It’s unclear where deeper geometric structure still matters at scale.

4. Geometry of training dynamics

Viewing training as a stochastic dynamical system:

* gradient descent as discretized flow * SGD as an SDE * trajectories on manifolds * attractors and metastability

This connects to dynamical systems, stochastic analysis, and geometric mechanics, but remains underdeveloped relative to its apparent relevance.

5. Discrete vs smooth geometry

Modern ML is deeply discrete: finite precision, quantization, sparse activations, graph-based computation. Smooth differential geometry may be the wrong limit in some regimes. Discrete differential geometry or combinatorial curvature might be more appropriate.

Some failures of “geometric ML” may simply be failures of choosing the wrong geometric category.

What I’m looking for:

* geometric structures that have actually influenced model or optimizer design beyond equivariance * where Riemannian or information-geometric ideas help in large-scale training * which geometric frameworks seem promising but currently mismatched with SGD * directions I’m missing

Perspectives from theory groups, applied math, and industry research labs would be very welcome.

Comments URL: https://news.ycombinator.com/item?id=46882060

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=46881100

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=45324363

Points: 5

# Comments: 0

[1] For every position you look for people at LinkedIn in that company. If you are already connected ask for referral.

[2] If not, look for common connections that can introduce you.

[3] If there is none => send request for adding everyone from the company to the friends.

[4] Message everyone, inviting for a coffee or virtual chat to learn about the company.

[5] If you believe that you are a good fit => say this.

----

And you spend first 4 hours of every day messaging, messaging, messaging. The response rate will be low. But you need only one job.

----

And if you do it for 6 months every day, there is no way you will not get many interviews.

----

Every interview you fail => you extensively study to address limitations of your skillset.

-----

Basically, there is no: "I have a degree, hence I deserve a job", but there is: "hard work is the goal".

-----

P.S. Somehow blog post reminds me why online dating is not working that well for men. Competition is enormous, the number of ladies is limited and things like: "I am an average man, hence I deserve attention from ladies" does not work as well

If we talk about popular packages: - people want it - people enjoy it - people do not pay for that

But force-feeding with strict licenses like Ultralytics does works. Yes, it is force-feeding, but noone wants to pay the price, unless there is no other choice.

Steps: [1] Add pre-commit hook [2] Write design doc as mdc file to .cursor/rules [3] Iterate on design doc till it describes what I want [4] Ask to write the code [5] Where possible - extend the test suite. [6] On every commit check that pre-commit hook checks and tests pass [7] On every bug extend the test suite [8] Write as many as possible custom pre-commit hooks [9] Add extensive docstrings to the complex code -> adds extra context to the LLM [10] Iterate [11] From time to time ask to verify that design doc is up to date

It would be better for design to be intuitive, but you struggle only the first time, while interfaces overloaded with information will take some of your attention every time you look at them

Makes training much more flexible, and fine tuning as well. Now, when a new data in terms of samples or new tables are connected to the olds ones you just extend the existing graph, without changing its existing morphology much.

Although it is unclear if it is scientific: "Look how cool we can do" or business result: "Look how much value do we get from this representation"

Useful by itself, but I suspect that will get even more adoption in industry in the telemetry as details statistics about users is ultra valuable.