I am soooooooooo fucking stupid!

I was using an np.empty instead of np.zeros to shape the Y tensor...
🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️🤦🏻‍♀️

No fucking wonder the poor thing was keep failing to detect a pattern.

Let's see if it works now... *play elevator music while we wait for keras*

As you can see from the screenshot, its working.
The system is actually learning the associations between the digit sequence of semiprime hidden variables and known variables.

Training loss and value loss are super high at the moment and I'm using an absurdly small training set (10k sequence pairs). I'm running on the assumption that there is a very strong correlation between the structures (and that it isn't just all ephemeral).

This initial run is just to see if training an machine learning model is a viable approach.

Won't know for a while. Training loss could get very low (thats a good thing, indicating actual learning), only for it to spike later on, and if it does, I won't know if the sample size is too small, or if I need to do more training, or if the problem is actually intractable.

If or when that happens I'll experiment with different configurations like batch sizes, and more epochs, as well as upping the training set incrementally.

Either case, once the initial model is trained, I need to test it on samples never seen before (products I want to factor) and see if it generates some or all of the digits needed for rapid factorization.

Even partial digits would be a success here.
And I expect to create multiple training sets for each semiprime product and its unknown internal variables versus deriable known variables. The intersections of the sets, and what digits they have in common might be the best shot available for factorizing very large numbers in this approach.

Regardless, once I see that the model works at the small scale, the next step will be to increase the scope of the training data, and begin building out the distributed training platform so I can cut down the training time on a larger model.

I also want to train on random products of very large primes, just for variety and see what happens with that. But everything appears to be working. Working way better than I expected.

The model is running and learning to factorize primes from the set of identities I've been exploring for the last three fucking years.

Feels like things are paying off finally.

Will post updates specifically to this rant as they come. Probably once a day.

Keras was throwing errors...

Since I thought it was a tensorflow issue, I went up and down and all the way around. Installing all tensorflow shit like a bijillion times.

... But it wasn't. It was the fucking ipykernel...
It took me a good 5-6 hours.

I pulled a 12 hours day today.

... Somebody hug me plz 😢

I've assembled enough computing power from the trash. Now I can start to build my own personal 'cloud'. Fuck I hate that word.
But I have a bunch of i7s, and i5s on hand, in towers. Next is just to network them, and setup some software to receive commands.

So far I've looked at Ray, and Dispy for distributed computation. If theres others that any of you are aware of, let me know. If you're familiar with any of these and know which one is the easier approach to get started with, I'd appreciate your input.

The goal is to get all these machines up and running, a cloud thats as dirt cheap as possible, and then train it on sequence prediction of the hidden variables derived from semiprimes. Right now the set is unretrievable, but theres a lot of heavily correlated known variables and so I'm hoping the network can derive better and more accurate insights than I can in a pinch.
Because any given semiprime has numerous (hundreds of known) identities which immediately yield both of its factors if say a certain constant or quotient is known (it isn't), knowing any *one* of them and the correct input, is equivalent to knowing the factors of p.

So I can set each machine to train and attempt to predict the unknown sequence for each particular identity.

Once the machines are setup and I've figured out which distributed library to use, the next step is to setup Keras, andtrain the model using say, all the semiprimes under one to ten million.

I'm also working on a new way of measuring information: autoregressive entropy. The idea is that the prevalence of small numbers when searching for patterns in sequences is largely ephemeral (theres no long term pattern) and AE allows us to put a number on the density of these patterns in a partial sequence, but its only an idea at the moment and I'm not sure what use it has.

Heres hoping the sequence prediction approach works.

Dude. Tensorflow version changes are so fucking bad. It's even worse with keras because they create an echo chamber for shit. I'm trynna reset a fuckin model here, yet everything throws 99 more errors to the pile. Like, wtf?

***** For stackoverflow enthusiasts: found a solution, don't need your groundbreaking shit either.

Keras went tits up...

┻━┻︵ \(°□°)/ ︵ ┻━┻

Reinforcement learning is going to be my end. 😩😩😩☠️

(currently stuck at how to put images as well as a bunch of other -motor- values as input... and exactly what am I getting as output again?)

Pulling my own hair out... Ooooooof

Just about every proper keras model I've ever made.

'Nuf said.

I can't be arsed with jobs that mention tensorflow alone as their main tech.

If your company is willing to use tf and not keras, then y'all probably didn't understand what you're dealing with to begin with.

*Red flags and sirens in distance for bad designs*

If you're an ml engineer, you must know how to hyperopt. I could recommend keras tuner tho, it's nice and saves shit on the go.

I get it — it’s not fashionable to be part of the overly enthusiastic, hype-drunk crowd of deep learning evangelists, who think import keras is the leap for every hurdle.

The fact that I managed to migrate the same fuckin keras model from gym to my own mujoco env and nothing broke too bad, is absolutely amazing.

Let's hope the little shit actually ends up learning some proper shit. 😒🦄

So I got the LSTM working in keras.
Working from a glorified tutorial.
Why the fuck do people let their github pages go down with no other backup?
Especially if its a link in your blog?
Why would you do that and not post the full script (instead of bits and pieces interspersed with *partial* explanations)?

In any case, its working and training on a test set and examples just to debug my own understanding of the process.

Once thats done I can generate some training data and try training on a small set. If that goes smoothly and the loss looks like it is heading in the right direction, then I'll setup the hardware for the private cloud and start writing the parallel computing component.

Hear me out:

Since keras and tf are pretty much schema design rn what if someone made a no-code solution where you drag and drop layers and tweak things in a UI so those data scientists can design it in a UI instead of writing shitty code?