jotaf

Planet Farms | Computer Vision and Machine Learning Engineer | Portugal | https://www.planetfarms.ag

Fully automated vertical farms.

We are a 50+ company with a working, prototype robotic vertical farm that is already shipping fully chemical-free vegetables farmed by robots to major European supermarkets.

Our international team of engineers, scientists and roboticists spans Lisbon, Milan and Oxford. As a CV/ML Engineer you will propose, implement and test new deep learning algorithms that will increase the autonomy of our vertical farm operations.

For more info, e-mail: joao.henriques at planetfarms dot ag

Planet Farms | Computer Vision and Machine Learning Engineer | Portugal or REMOTE | https://www.planetfarms.ag

Fully automated vertical farms.

We are a 50+ company with a working, prototype robotic vertical farm that is already shipping fully chemical-free vegetables farmed by robots to major European supermarkets.

Our international team of engineers, scientists and roboticists spans Lisbon, Milan and Oxford. As a CV/ML Engineer you will propose, implement and test new deep learning algorithms that will increase the autonomy of our vertical farm operations.

For more info, e-mail: joao.henriques at planetfarms dot ag

Thanks, but I think that lucb1e's confusion was probably the same as mine -- given pretrained CLIP features, how is this translated to zero-shot tracking?

Are initial bounding boxes given as usual, or are objects of interest created automagically?

Or are they tracked just from text descriptions?

Lots of questions after reading the post :)

There's also Siamese Networks, they're pretty fast and simple, and people keep updating their backbone CNNs to more modern (computationally expensive) ones.

significant amount of engineering cost to continue to deliver static sites, even with reduced fidelity and a worse user experience

A few things to consider:

- The reputational damage of solidifying the narrative that Google drops projects without a good reason from the users' point of view.

- A bare-bones deletion/flagging UI does not seem like a huge undertaking (though as an engineer I realizing it's not exciting work); working things out with a non-profit such as the Internet Archive or a museum could be an even lower-cost solution.

- 90's websites are also hideous ("worse user experience") by today's standards, yet they have charm and are a product of their times. I don't think anyone would argue that they should be deleted on account of that, just like ancient pottery often has "badly drawn" human figures yet the value is in the cultural expression. There are digital conservation efforts by museums (e.g. restoring 80's arcade games); I'm sure that consulting with a digital conservator would have arrived at a very different treatment of this data.

Good job, this looks extremely expressive and with fewer corner cases ("dark knowledge") than MPL/SNS, and unlike Altair/Plotly doesn't require a whole browser to display the output!

Still, I'd like to ask if you considered alternatives to MPL for the back-end. It's a venerable but ancient project with years of accumulated technical debt, and I'm sure you had to deal with lots of inconsistencies there.

For example, PyQtGraph is an alternative with a clear class hierarchy and can handle large-scale datasets without slowing down (while anything non-trivial in MPL has you wait seconds to render).

(I'd love to hear more suggestions that don't require a JS engine and don't build on MPL.)

It depends what you call the "home row" and where you keep your fingers.

If you expect to use these shortcuts (typically with the outer fingers of your hands) your hands will rest at a slightly higher position in your keyboard, with the pinkies near Esc and Del, and the thumbs near the spacebar. (Other fingers over the letters.)

I guess it depends on what you've trained yourself to do.

Just to play devil's advocate, these 3 examples don't seem much faster than Shift+(Del,Del,Del), Ctrl+F,string, and F3, respectively.

Should've updated the title! Now it converts HTTP requests to proper 3D meshes (seems to work again).

You have tab completion in the search bar (Ctrl+K)!

Personally I disable the cloud-based suggestions and rely only on my local history.

Not affiliated in any way, but you should give Signal a try.

The hardest part with any IM service is going to be convincing your friends to join though. I wish someone came up with a good way to overcome this "critical mass" problem.

Hi, I wasn't very happy with the existing options for managing/visualizing deep learning experiments, so (naturally) I rolled my own (and spent way too much time in the process).

Specifically, what I didn't like was: 1) The strange ritual of launching a terminal command and being told to open my browser in a local port always struck me as a bit clunky. 2) I want to visualize tensors and do arbitrary plots as I go, with MatPlotLib. (TensorBoard specifically can be restrictive in the plot types.)

So here's the result; a fully responsive app in PyQt. There's a GIF showcasing most features after the link!

When I read this my first thought was "you should use a FrozenDict", but looking into it, it seems that it doesn't exist; I assumed so because of FrozenSets (hashable sets).

You can still create a hashable key from a dict with FrozenSet(mydict.items()).

These are my favorite lecture notes, they assume almost no a-priori knowledge (with an awesome review of basic probabilities) and yet they don't shy away from explaining all the rigorous math.

If you have time to read step-by-step derivations and want to understand the fundamentals, I think this is an excellent self-contained resource.

https://ermongroup.github.io/cs228-notes/