Wizards Den

Bad news: I'll have to start over
Good news: The libs will speed up everything and shorten down my code significantly, so I can still be done tonight ^^ Assuming I don't screw up

ProgramFOX

5:16 PM

be done tonight, meaning that we can already test it here tonight?

Princess Luna

Good question

This really depends on how fast I finish vs training speed

I'm gonna be running tests locally and if I get stuff like "thethethe" as the output, it's not gonna be up for a while

but yeah, it could

ProgramFOX

"thethethe" > no output at all

Princess Luna

I'll have to fix the frontend again though

ProgramFOX

Besides, "thethethe" probably makes more sense than 50% of @Sha's messages.

2

Princess Luna

XD

@sha :D xD

Training data ready:

meta, idx_q, idx_a = data.load_data("")
(trainX, trainY), (testX, testY), (validX, validY) = data.split_dataset(idx_q, idx_a)

trainX = tl.prepro.remove_pad_sequences(trainX.tolist())
trainY = tl.prepro.remove_pad_sequences(trainY.tolist())
testX  = tl.prepro.remove_pad_sequences(testX.tolist())
testY  = tl.prepro.remove_pad_sequences(testY.toList())
validX = tl.prepro.remove_pad_sequences(validX.toList())
validY = tl.prepro.remove_pad_sequences(validY.toList())

xLen = len(trainX)
yLen = len(trainY)

Princess Luna

5:44 PM

dataset spliced and diced in 10 seconds xD

the same thing I spent 9 hours on yesterday xD

ProgramFOX

hah, wow

KennyBOT

6:47 PM

~ The Wizard of Oz

Princess Luna

and I'm back from dinner. Everything appears to be working so far at least, data loading works perfectly

ProgramFOX

7:15 PM

great!

Princess Luna

Hell yes! NLP data allows integration with intent, meaning I can add intents like "what's your name", "tell me a joke", etc, that allows me to actually create external memory modules

If I set it up properly, I can load in a bunch of predefined classes and make the bot find examples itself

I can throw in an NLP network for grammar correction and fix capitalization, reducing the amount of words the bot needs to know (because someword != Someword)

Btw, with the current level of motivation and typing speed, the network should be finished in 20 minutes, I'll finish up the bot and frontend while it trains and I can probably get it up by 11

Princess Luna

7:48 PM

so what, 3 nets total? xD

for a SE/Discord/console chatbot xD

#overkill

Shadow Wizard

@PrincessLuna great!

KennyBOT

~ Stars get removed under peer-pressure?

Shadow Wizard

@ProgramFOX jhghugyfg

FOX 9000

No associated word found for jhghugyfg.

ProgramFOX

point proven

Shadow Wizard

7:58 PM

ךםך

Princess Luna

tl.files.load_and_assign_npz(sess=sess, name="n.npz", network=net)

That line right there loads the saved data for the network that goes into training in 30 seconds

And here we go

For science! :D

Mooseman

8:16 PM

@PrincessLuna Yesh

@Moosebot salute

user280395

@Mooseman SIR, YES SIR! 21-gun salute now commencing...

Princess Luna

@Moosebot kill @Moosema

user280395

@PrincessLuna Nice try. They're not on my hit list... yet.

user280395

> @Moosema was killed by Col. Mustard with the Wrench in the Billiard Room.

user280395

> @Moosema was crushed when the TARDIS landed on their head.

Princess Luna

8:17 PM

:>

Hiya @Mooseman!

Princess Luna

8:28 PM

Input:  great
>  i
>  end_id you
>  end_id
>  you you you
>  end_id

Great results :P

there's no sentiment analysis though :/

ProgramFOX

8:49 PM

Hmm, very useful indeed :P

Princess Luna

Training is slow though. The second epoch is finished(/50) so it's not gonna be tonight

ProgramFOX

ah

Princess Luna

Input:  great
>  i dont have to a unk
>  you dont to to to to a unk
>  you dont have to a unk
>  you dont to to a unk
>  you have to a unk

dammit

I need NLP

ProgramFOX

what's a unk?

Princess Luna

unknown. It's in the list of words to avoid issues with words not present in the vocab

ProgramFOX

9:03 PM

ah

Princess Luna

9:33 PM

the inevitable problem is training data. That's where I now too see I get stuck

once I get it rolled out to the SE network, I can start collecting data from the network and train as it goes along. The problem is figuring out the vocab though, since when it increases it consumes more memory

if necessary I can decrease the input dim to fit the amount of words though, but the training data alone should correct some issues. But there's also the personality which I think I have to use NLP for in order to find cases where memory matters

alternatively, I could try replacing all the names in output with the bot name, but that's a ton of data scrubbing

Princess Luna

10:01 PM

Anyone else watching the nasa launch?

Jenna Sloan

@PrincessLuna NASA doesn't launch shuttles anymore

Princess Luna

They still launch stuff

Jenna Sloan

Just because a rocket launched from their launch site doesn't mean that the rocket is theirs

NASA is just nice enough to let the United Launch Alliance rocket launch from their Air Force Station

btw, here's the vid: