Smart India Hackathon: Horrible experience
Background:- Our task was to do load forecasting for a given area. Hourly energy consumption data for past 5 years was given to us.

One government official asks the following questions:-
1. Why are you using deep learning for the project? Why are you not doing data analysis?
2. Which neural network "algorithm" you are using? He wanted to ask which model we are using, but he didn't have a single clue about Neural Networks.

3. Why are you using libraries? Why not your own code?

Here comes the biggest one,

4. Why haven't you developed your own "algorithm" (again, he meant model)? All you have done is used sone library. Where is "novelty" in your project?

I just want to say that if you don't know anything about ML/AI, then don't comment anything about it. And worst thing was, he was not ready to accept the fact that for capturing temporal dependencies where underlying probability distribution ia unknown, deep learning performs much better than traditional data analysis techniques.
After hearing his first question, second one was not a surprise for us. We were expecting something like that. For a few moments, we were speechless. Then one of us started by showing neural network architecture. But after some time, he rudely repeated the same question, "where is the algorithm". We told him every fucking thing used in the project, ranging from RMSprop optimizer to Backpropagation through time algorithm to mean squared loss error function.

Then very calmly, he asked third question, why are you using libraries? That moron wanted us to write a whole fucking optimized library. We were speechless at this question. Finally, one of us told him the "obvious" answer. We were completely demotivated. But it didnt end here. The real question was waiting. At the end, after listening to all of us, he dropped the final bomb, WHY HAVE YOU USED A NEURAL NETWORK "ALGORITHM" WHICH HAS ALREADY BEEN IMPLEMENTED? WHY DIDN'T YOU MAKE YOU OWN "ALGORITHM"? We again stated the obvious answer that it takes atleast an year or two of continuous hardwork to develop a state of art algorithm, that too when gou build it on top of some existing "algorithm". After listening to this, he left. His final response was "Try to make a new "algorithm"".

Needless to say, we were completely demotivated after this evaluation. We all had worked too hard for this. And we had ability to explain each and every part of the project intuitively and mathematically, but he was not even ready to listen.
Now, all of us are sitting aimlessly, waiting for Hackathon to end.😢😢😢😢😢

So I have a teacher that when he use "C++" it is basically C with a .cpp file-extension and -O0 compiler flag.

Last assignment was to implement some arbitrary lengthy calculation with a tight requirement of max 1 second runtime, to force us to basically handroll C code without using std and any form of abstraction. But because the language didn’t freeze in time 1998, there is a little keyword named "constexpr" that folded all my classes, arrays, iterators, virtual methods, std::algorithms etc, into a single return statement. Thus making my code the fastest submitted.

Lesson of the story, use the language to the fullest and always turn on the damn optimizer

Ok now I’m done 😚

Computer fixer, software cracker, smartphone optimizer, and a hacker just because I use terminal.

At least they love me.

Traffic lights optimizer. Imagine big AI being plugged to all city's traffic lights and changing all durations in order to adjust all of them, if we add smart cars connected to it we might remove traffics at all. It is way ahead in future but would be cool.

so after several hours of irritated detective work, I've finally found out what is the thing that periodically, every about 10-15 seconds, starts two PowerShell processes which run for about a second or two and during that time take about 20% of my CPU capacity...

They're being launched from a commandline, to do GetPackages with name of OmenLightStudio, and the result is then piped into find.exe to find InstallLocation part.

...for whatever reason.
and this is done every 10 seconds by... *drumroll*
HP SYSTEM OPTIMIZER.

GOD. FUCKING. DAMMIT. YOU. MORONS.

...now only to find how the fuck do I uninstall that, since it's some plugin-ish kind of stuff for Omen Studio, and I can't find uninstall for it anywhere in the system nor Omen Studio itself...

OK so... project I've been working on! It's a virtual processor that runs in the browser coded in JavaScript. OK so I know, I know, you must be thinking, "this is crazy!" "Why would she do this?!?!" and I understand that.

The idea of Tangible is is to see if I can get any tangible performance over JavaScript. I've posted a poorly drawn diagram below showing how tangible works.

The goal for tangible is to not use html, javascript, or CSS. Instead, you would use, say for instance, c++ and write your web page in that, then you compile it using my clang plugins and out pops your bytecode for Tangible. No more CSS, no more html, and no more javascript. Instead everything from a textbox to a video on your web page is an object, each object can be placed into a container, each container follows specific flag rules like: centerHorizontal or centerVertical.

Added to all of this you get the optimization of the llvm optimizer.

I hate iOS so much. Just hit a nasty bug where "optimizer" in browser sometimes randomly replaces a function with a number. Like WTF? Was developing a webpage working on iOS too easy, so they enabled a hard mode where some of your functions may turn on you?

Today I had a full-day job interview for a junior data scientist position.

First I met the team which was only like half of everyone because apparently everyone was gone on Fridays. However the few there were really nice.

First task is to do some basic data analysis stuff even though I already spent a week on the coding challenge and sent them all my code/tasks. I log into my machine and create a new virtual environment but can't for the life of me figure out how to use the command line in windows to install packages. Turns out there is some problem with their proxy and they have to log me in on that. Then I am struggling on the keyboard because it's for a language different that my mother tongue and it takes me 3x as long to so the most simple things. All my shortcuts are out the window. Haven't a hard time typing parentheses and brackets. Start freaking out and have a panic attack mid task. I'm sweating bullets. I didn't even make it to the simple visualization tasks much less the models at the end. Time gets called and we all go to lunch and I'm freaking out on the inside the entire time. Angry at myself because I know I am better and just couldn't think.

After lunch I present my code and results from a coding challenge I did weeks prior. People from other teams get invited and I end up getting grilled for 2 hours by 15 people. Questions are flying in from all sides. They ask me almost everything I know about machine learning and some more. Under stress I forgot the name of the optimizer I used and couldn't answer some easy stuff because my mind was racing.

Right now I am on the train home and my body physically hurts. I am disappointed with myself and wish I could have shown up better. Never really froze up like this before.

I finally got the lstm to a training and validation loss of < 0.05 for predicting the digits of a semiprime's factors.

I used selu activation with lecun normal initialization on a dense decoder, and compiled the model with Adam as the optimizer using mean squared error.

Selu is self-normalizing, meaning it tends to mean 0 and preserves a standard deviation of one, so it eliminates the exploding/vanishing gradient problem. And I can get away with this specifically because selu *only* works on dense layers.

I chose Adam, even though this isn't a spare problem, because Adam excels on noisy problems and non-stationary objectives (definitely this), and because adam typically doesn't require a lot of hyperparameter tuning its ideal here, especially considering because I don't know what the hyperparameters should be to begin with.

I did work out some general guidelines on training quantity vs validation, etc.
The initial set wasn't huge or anything, roughly 110k pairs for training.

It converged pretty quick all things considered, and to the low loss like I mentioned, but even then the system always outputs the same result, regardless of the input, so obviously I'm doing something incorrectly.

The effectiveness of this approach for training and validation makes me question if I haven't got something wildly wrong. Still exploring though and figuring out how to get my answers back out. I'm hoping I just fucked up the output, and not the input as well.

- C# call to SQL Server takes forever.
- Running the same sql in SSMS is nearly instant.

Please SQL Server God, grant me strength to understand your ways.

There’s somewhat of a magical moment when you teach the interns how to turn their huge switch statement into a function Pinter branch table....

Eliminating shotgun surgery as the project advances and new features are added to a module.

Colleges definitely don’t teach students certain things... and when I end up teaching them... the excitement and ideas they come up with are always great.. and they always end up with a stronger understanding of things.

(Embedded company .. all software here is C or C++) yes we use the optimizer.... sometimes you gotta do seemingly complex things to improve readability and maintainability.

SQL is amazing.

I'll toss out some bassakwards query and the optimizer will make sense of it and suddenly I'm searching a amazonillian records in no time.

Then rando one day (today) I fire up what I think is really not the most wonky query I've ever written and ... "Well shit this is surprisingly slow."

So then I go full n00b and add some fields to the query that I know would limit the number of possible records to way low thinking that might help and ... nope no faster...

Guess it's time to bust open some books about SQL....

ohhhhh I am pissseddddddddddd

itss the fucking pytorch.module class it would seem !

I do exactly the same goddamn shit as its supposed to do in a goddamn notebook and run it step by step and the fucking model trains and the output values change !!!! and the loss decreases !!

I do this in the goddamn class derived from model with a call to model.parameters() and the fucker fails !!!

why ???
why ?????
why ??????
is it cloning the goddamn parameters so the references aren't there ????
seems to work goddamn fine when i call a layer and activation function at a goddamn time chaining the calls one after another !!!!!!

UGHHHH IT LOOKS LIKE IF YOU DEFINE THE LOSS AND OPTIMIZER OUTSIDE THE FUCKING CLASS IN A SEPERATE TRAINING FUNCTION IT DOESN'T TRAIN !!!!!!

WHY ??
A REFERENCE IS A GODDAMN REFERENCE !!!!