eggie5

City taco in north park

I lost my sense of smell and taste after a head injury. I remember walking into a Starbucks and it wreaking of fish! The hardest part was when I first bit into a cookie, post injury, and I couldn't recognize the taste :(

Took months to get it all back.

I like to use Typora w/ iCloud

how did you refinance so low?

I have really flat feet and as you know the arch on birks is vey aggressive. so it just rubs on my foot super hard. If I know I'll be walking far I'll put athletic tape on my non-arch or else a blister will form. Now I have visible calluses build up, but I persist... :(

all the good ML tools are in python

I would say tensorflow is the anthesis of "convention over configuration" rails is lauded for! w/ TF relative to most frameworks, you have to configure all the boilerplate!

If you're looking for a higher-level approach a lot of the community likes sci-kit which has lots of nice defaults for popular models.

ok, I see, you are making the connection on basis of hybrid characteristics.

Since you're familiar w/ the youtube paper, I've been wondering this question: How do they get vectors out of the softmax?

so you save your model weights to Postgres, expose the model via an HTTP api using flask and package it w/ docker?

can you please elaborate?

How would you generalize the method you are using?

I don't understand your connection between lightfm and the youtube paper...

lightFM is my goto for prototyping Matrix Factorization models. It efficiently handles large data w/ sparse data structures and is CPU accelerated including optimizations like Hogwild!. It also has the WARP loss BPR variant which I have not seen implemented anywhere else.

I can train on multi-GB datasets w/ only lightFM and multiple CPUs.

Another interesting package is called Implicit. This package, although not as complete as LightFM when it comes to algorithms or APIs, really shines when it comes down to optimizations. Including native Cuda kernels for BPR and ALS it also has an important speedup called the Conjugate Gradient Method which makes it faster than spark in some benchmarks.

But usually, now-a-days my work requires more customized hybrid models of which I usually start w/ a base BPR implementation I have in Keras.

Trivago has had success

obike shut down? I remember they were all over Amsterdam last year

The cool thing about the ICE trains and NS trains (Netherlands) is that you can get to the station and just walk on the train. Unlike in Spain where there are long security X-ray lines.

rails

Heroku free dynos

I make way more money in the US but my quality of life was much better in Netherlands

a big problem w/ recommender systems is called the cold-start problem. One way this is manifest is when new items enter the system. In one class of recess, collaborative filtering, since the new item doesn't have any interaction history, it will never be recommended. Then in your case where news articles are the items, then you have an extreme case where every single item is cold.

Look at a talk from a Dutch group called Blendle at RecSys conf where they talk about these problems.

"my first quick experiment was a patch based "bee / no bee in image" detector. "

What are these patches and what was your intuition on starting w/ it?

they posted on Who's Hiring back on November last year

also useful for incorporating noisy GPS sensor readings! I tried to do this in an undergrad autonomous vehicle but failed at the time :(

I try to do as much of the transformations as possible in tensorflow using the Datasets API so that I don't have to write them in another language/system in production.

here's my attempt a while back: http://www.eggie5.com/66-naive-bayes-classifier-implementati...

a lot of people using spark?

This sounds like the classic image classification challenge. The next question I would have is: is this on open world or closed world problem? Meaning, what is the world of images that the model will be exposed to in the wild. Will it only be exposed to sailboat/windmill/bird or will it be exposed to sailboat/windmill/bird and every other possible item in the world?

You also want to deploy this on a standard machine w/ or w/o a GPU? That can have implications on inference speed. But there are ways to optimise for that. The absence of a network connection is not a problem.

Also, do you have training data? If not we can probably leverage a pre-trained model for this use case where, in this case we would only need a handful of training examples.

To answer some of your questions:

Where should I look to find a quality freelancer? Not sure.

What formats should I specify for the deliverables? Depends on what tools the developer uses, for example, I would deliver files in TensorFlow export format.

What timeline should I expect? Using a pre-trained model, I could expect to deliver it in about 40 man hours.

What pricing should I expect? Probably about $6000, but that depends on the contractor.

What would a good developer expect me to provide in terms and training data? Using a pre-trained model you would only need a handful of training images. But, of course, the more the better.

What API parameters would they expect me to specify? I can't think of anything.

I have experience building and deploying these vision models. Feel free to reach out to me for more info. It's my username @ gmail.com

2 jobs

I realised there is a lot of variance in the process and only a small amount of the variance is under my control. So much is outside factors due to random chance. Therefore, I decided that since I have little control over the outcome, then the more effort I put into an application, the more energy/time I waste. So instead I decided to optimise for quantity and not quality. The way I scale it up is to automatically apply to everything on who's hiring monthly posts that have an email address. I have a std email template that I have honed over the months. Otherwise I can apply on a website in under 45 seconds. Every month I'll spend 20 minutes get out at least 30 applications of which 80% will convert to recruiter calls of which about 95% will convert to tech screens of which %10 (I suck at coding puzzles under pressure) will convert to onsites of which 15% will convert to offers.

As long as you diligently use your calendar and keep notes for each company it's pretty easy to manage this.

I can vouch for the leadership at this company!