The WPCC Lounge

[Ganaram Inukshuk 3,682]

Just now, CypherHoof said:

Well, I am glad you are interested in Data Science (we need more DS in the world) but my primary focus is on cybersecurity (which has some DS and some ML) but for me, DS is more a means to an end not a primary goal.

Finally someone who knows something about ML... (I've been wanting to use a genetic algorithm for a video game, of all places, when the last time I ever trained a machine was to have it try to write a FiMfic based on the longest stories on Fimfiction that I could find... Also Markov chains, which I used for a class project...)

[CypherHoof 26,483]

2 minutes ago, Longhaul said:

Actually, that might be what I need to get me through some of these long nights.

Well, Strong Drink is a mocker, but a Coffee and Chocolate mix is a Mocha

2 minutes ago, Ganaram Inukshuk said:

Finally someone who knows something about ML... (I've been wanting to use a genetic algorithm for a video game, of all places, when the last time I ever trained a machine was to have it try to write a FiMfic based on the longest stories on Fimfiction that I could find... Also Markov chains, which I used for a class project...)

I tend to find that ML generated stories tend to throw in a "wait, what?" moment after every third sentence or so; they look superficially convincing (in that any individual sentence or pair of sentences seem plausible) but lack of an overall plot line or any real continuity seems to break the output. No doubt that will be fixed (or at least, some hardcoding of a Universal Plot and so forth should help).  Markov chains are part of the problem there (and cause the "wait, what?" moments) due to lack of dependency; new output is not constrained to continuity with past output.

Amusingly, we were using Markov chains in computing for years (as finite state machines) before we started calling them that :D)

Surprising you don't find more ML knowledge amongst the herd though; compute-focussed ponies seem pretty common on MLPF.

[Ganaram Inukshuk 3,682]

2 minutes ago, CypherHoof said:

Well, Strong Drink is a mocker, but a Coffee and Chocolate mix is a Mocha

I tend to find that ML generated stories tend to throw in a "wait, what?" moment after every third sentence or so; they look superficially convincing (in that any individual sentence or pair of sentences seem plausible) but lack of an overall plot line or any real continuity seems to break the output. No doubt that will be fixed (or at least, some hardcoding of a Universal Plot and so forth should help).  Markov chains are part of the problem there (and cause the "wait, what?" moments) due to lack of dependency; new output is not constrained to continuity with past output.

Amusingly, we were using Markov chains in computing for years (as finite state machines) before we started calling them that :D)

Surprising you don't find more ML knowledge amongst the herd though; compute-focussed ponies seem pretty common on MLPF.

To quote someone trying to get a neural network to dance (technically two NN's), we need a bigger brain. (And also more training data.)

I'm starting small, like "getting a car to drive around in Unity" small, but I don't want each car to drive perfectly and with the same brain (unless I'm talking simulating self-driving cars), but instead I want variation (because most of the genetic algorithms in Unity vids only ever pit one agent against the environment, not multiple agents trying to avoid each other while also traversing said environment).

It was just last year that I started writing a Markov Text Generator for generating text (and I still have it), and creating a user interface for it became an actual school project (on top of porting it over to Javascript).

I have too many secondary interests. Also, this:

Surprising you don't find that ML knowledge amongst the problem there (and cause the problem there (and cause the output. No doubt that ML generated stories tend to find that any real continuity seems to throw in a Universal Plot and so forth should help).á Markov chains are part of dependency; new output is not constrained to lack of dependency; new output is not constrained to lack of an overall plot line or pair of dependency; new output is not constrained to find more ML generated stories tend to throw in a Universal Plot and so forth should help).á Markov chains in computing for years (as finite state machines) before we were using Markov chains are part of a Universal Plot and so forth should help).á Markov chains in a Universal Plot and so forth should help).á Markov chains in a "wait, what?" moments) due to continuity with past output.

I also have it so that my MTG can print the entire state table; my Javascript version can do that in JSON format.

[Ganaram Inukshuk 3,682]

Actually, it gets worse. Markov Music Generator except the input data is inherently two-dimensional. I also don't have a way of converting sheet music into something one-dimensional, and also a piano has 2^88 unique states (any moment in time in which the keys are pressed down, including all 88 keys pressed down at once and none of them pressed down; foot pedals just add 8 times more states).

To reduce the number of states to account for, you'd have to slap a bunch of restrictions, such as "no more than, say, 5 notes played at once" and "one key only" and "but every song is transposed to the key of C" and a bunch of restrictions on how many octaves you're allowed.

[CypherHoof 26,483]

3 minutes ago, Ganaram Inukshuk said:

Actually, it gets worse. Markov Music Generator except the input data is inherently two-dimensional. I also don't have a way of converting sheet music into something one-dimensional, and also a piano has 2^88 unique states (any moment in time in which the keys are pressed down, including all 88 keys pressed down at once and none of them pressed down; foot pedals just add 8 times more states).

To reduce the number of states to account for, you'd have to slap a bunch of restrictions, such as "no more than, say, 5 notes played at once" and "one key only" and "but every song is transposed to the key of C" and a bunch of restrictions on how many octaves you're allowed.

Looking at midi streams might be instructional then. midi abstracts the processing of sheet music, almost literally (in its simplest form, a midi stream and actual notes on staves are interchangable, and many music packages such as sibelius, finale, lillypond and so forth can load and save stave information as midi files).  As midi files (as dot-mid) are pretty common and available for most classical pieces, getting a decent data pool for ML isn't hard.

Haven't really seen much lately on music generation, but I remember a lot of stories about it a year or so ago, so maybe you could start by looking at published research in the area, rather than reinventing the wheel there?

[Ganaram Inukshuk 3,682]

Just now, CypherHoof said:

Looking at midi streams might be instructional then. midi abstracts the processing of sheet music, almost literally (in its simplest form, a midi stream and actual notes on staves are interchangable, and many music packages such as sibelius, finale, lillypond and so forth can load and save stave information as midi files).  As midi files (as dot-mid) are pretty common and available for most classical pieces, getting a decent data pool for ML isn't hard.

 Haven't really seen much lately on music generation, but I remember a lot of stories about it a year or so ago, so maybe you could start by looking at published research in the area, rather than reinventing the wheel there?

Based on what little I know about music-- actually, I know very little about music, only that you can turn it into a massive state table or directed graph and the arrangement of keys of your standard piano requires you to learn the piano twelve times when ideally you should only have to learn it once (isomorphic alternatives exist but are rare).

I can only ever refer to one example where someone took and processed a whole bunch of MIDI files (first MIDI files of baroque music, then jazz), mapped the 88 keys of a piano represented in MIDI to 88 unique ASCII non-whitespace characters, and fed the resulting text into a neural network originally designed to generate text. He also explained some of the issues involved with generating music, including how music is inherently two dimensional and how a transposed version of a song is treated as a whole new song. (He later resorted to a GAN and then a CNN to generate jazz music, which entailed translating MIDI files to image files, feeding that into the GAN/CNN, having the GAN/CNN generate new images, and translating those images back into MIDI files.)

The RNN was Andrej Karpathy's Character RNN; the person who attempted all of this music generation was carykh.

[Rixton 1,105]

On 12/28/2018 at 9:03 AM, Alexshy said:

Still busy, but I shall be free as the wind on morrow after the lunch and till January 9th.

Sounds good, more than a week of relax is a good start of the year.

59 minutes ago, CypherHoof said:

Looking at midi streams might be instructional then. midi abstracts the processing of sheet music, almost literally (in its simplest form, a midi stream and actual notes on staves are interchangable, and many music packages such as sibelius, finale, lillypond and so forth can load and save stave information as midi files).  As midi files (as dot-mid) are pretty common and available for most classical pieces, getting a decent data pool for ML isn't hard.

Haven't really seen much lately on music generation, but I remember a lot of stories about it a year or so ago, so maybe you could start by looking at published research in the area, rather than reinventing the wheel there?

I remember back when I was working on to play 1 tone music on a micro controller.

Anyway, how is everypony today?
Here it is a little bit foggy weather, not an energy full tho, I better like studying on a more clear day, maybe a cup of coffee would help.

[CypherHoof 26,483]

41 minutes ago, Rixton said:

Sounds good, more than a week of relax is a good start of the year.

 

I remember back when I was working on to play 1 tone music on a micro controller.

yeah. I remember trying to translate sheet music into frequency+duration for eight bit

 

41 minutes ago, Rixton said:

Anyway, how is everypony today?

Seem to have picked up a bug - cough, blocked nose, that sort of thing. hopefully will get over it during the weekend.

41 minutes ago, Rixton said:

Here it is a little bit foggy weather, not an energy full tho, I better like studying on a more clear day, maybe a cup of coffee would help.

I am sure there are problems for which a cup of coffee is NOT the answer. but this is the WPCC - we don't CARE about those problems

*goes and fires up the espresso machine for Rixton

[Rixton 1,105]

12 minutes ago, CypherHoof said:

Seem to have picked up a bug - cough, blocked nose, that sort of thing. hopefully will get over it during the weekend.

In my country there is a joke: if pálinka is not a cure for something, than there is no cure for that.

12 minutes ago, CypherHoof said:

*goes and fires up the espresso machine for Rixton

*Accepts the coffee. Thank you very much!

[Longhaul 8,441]

*enters in search of coffee*

[Flutterstep 47,324]

*passes over a fresh cup of hot coffee for @Longhaul*

[Twilight Luna 31,492]

So, I was in my local Target today and all their MLP toys were marked for clearance and Littlest Pet Shop toys were put in their spots. Looks like Target won’t be carrying MLP stuff anymore. 

Edited December 29, 2018 by Twilight Luna

[Flutterstep 47,324]

Maybe one of the brony employees took them all off the shelves for themselves.

[Twilight Luna 31,492]

Could be. I found them sitting in the clearance section.

[Longhaul 8,441]

There's a couple of Targets near me and both of them seem to have removed the MLP section.

[Ganaram Inukshuk 3,682]

Anyone else with a Blender (the program, not the machine that turns fruit into drinks)...?

[Twilight Luna 31,492]

44 minutes ago, Longhaul said:

There's a couple of Targets near me and both of them seem to have removed the MLP section.

Target didn’t want to invest in any of the new merchandise I guess. It’s a shame Hasbro didn’t promote season 8 toys more heavily. 

[CypherHoof 26,483]

1 hour ago, Ganaram Inukshuk said:

Anyone else with a Blender (the program, not the machine that turns fruit into drinks)...?

Got a copy, but sadly not very artistic. Why?

[Ganaram Inukshuk 3,682]

6 minutes ago, CypherHoof said:

Got a copy, but sadly not very artistic. Why?

Umm... Nothing...

[CypherHoof 26,483]

40 minutes ago, Ganaram Inukshuk said:

Umm... Nothing...

That is very well done, I am assuming Trixie? -

wheels seem to have a very low polygon count though?

[Ganaram Inukshuk 3,682]

2 minutes ago, CypherHoof said:

That is very well done, I am assuming Trixie? -

wheels seem to have a very low polygon count though?

The style's kinda the point and yes Trixie.

Spoiler

 

[Rixton 1,105]

Noice Blender scene!
I maybe arrived late, but what would be your question? @Ganaram Inukshuk

Edited December 30, 2018 by Rixton

[Ganaram Inukshuk 3,682]

5 minutes ago, Rixton said:

Noice Blender scene!
I maybe arrived late, but what would be your question? @Ganaram Inukshuk

There wasn't really a question, I was seeing if anyone else knew how to use it...

Also... 

(This pic is like several weeks outdated already; I've since refined the models and programmed the car to be moved using user input.)

[Rixton 1,105]

2 minutes ago, Ganaram Inukshuk said:

There wasn't really a question, I was seeing if anyone else knew how to use it...

Also... 

 

(This pic is like several weeks outdated already; I've since refined the models and programmed the car to be moved using user input.)

Ah, Unity! I wanted to get it started years ago, but than realized, I am too dummy for this.

[CypherHoof 26,483]

12 hours ago, Rixton said:

Ah, Unity! I wanted to get it started years ago, but than realized, I am too dummy for this.

Just realised I have been playing StarLink for twelve hours straight - so taking a break