Friday, October 21, 2016

Training Epoch 32 of 50 - Two Poems

This is a sample generated from training epoch 32 of 50 with middle-ground temperature.

mine with


   it was into the unknown bars
  to be me
  then he said, 
 I cant live
  who stands and sit my woman and
  somebody said, we walk to
  space that we are the words down at the eyes
  with the song of them
  being a child woman
  and the arms of the forest.
  the most old ass the large can be
  the houses and the bartender the streets,
  the silver in the least dispace
  and the raining to see the singtery
  white day, she said, man.
  I get her air. I could be she said
  as the mountains and asses and he was just like the sillieve hills
  and the other had pressed and have to die
  with the girls and her broken
  bed in a ceul walked off and
  an absent the bed with a whore
 
 And here is one with a higher temperature, which still produces quite a few errors. 
 
 or he can boil


  it saw, Bobing Man, yellow-sta-sceece new out
 legs and it in a stick for no injagent, let it is. lived at
 the brokenmay easy entart and packanceing win
 21! he said.
 it was the dog-on by my car from my Time
 on Harry.
 I listening to a mentel, I thought content
 good.
 I settle kits open in me think any over in her bar bowed there's
 and some hands a mit;
 Teddy.
 its all right, I did 
        beautifuating.
 

Thursday, October 20, 2016

Midway through Training with the new Files

We're half way though training. Here's a newly generated poem.


     Who didn't seem for the person from her
         Or her burn and went on the sun of the special house;
         And the most as would start to record when the hours
     If might used the soul,
     And as he's beat but his corner from the world;
     When casked with some sure of them have that your bedroom ready,
                          To sweet of earth
                          - where I was the other must here they sleep,
         With his short beer in the reason;
     The small the best boy song,
         And for the stairs that the same books
          And for it will we are the words I had a cross thoughts of the sun
         As some good for the raper
            But the line and his friends are beyond
       That thought that so meety was the light.







Tuesday, October 18, 2016

12th epic (of 50) Training Update



The training is progressing much quicker with the updated training file.

his ending
    the door.
    anyhow the best books street
  and catch of cutting not many there and she said,
  drunk.
    a time I remember some floor
  sits the prietless for a
  room
  the bumble off from the heat
  me.
    anyhow, the peace of us
  and then
  but she had any
  life wine
  and all your painting
  but what
  has to write a place
  sunfing your car steps
  on the patrouch where its
  as all the way manica room
  the tires of
  the window.
    that is there were is a freakt to
  my car.
    when Ive know what it was trying to know
  with men walking the light innor
  love thrives.
  this stages and fact 
 
There is a bit of context showing up now, car stuff, tires and windows and the repeated anyhow.  

Sunday, October 16, 2016

Reset in training after discovery of a corrupt training file

After a setback with a problem in the training file, we're back on track. We noticed some weird results were coming through and it didn't seem to be learning as well as it should. Turns out part of the file was corrupt.

So we're back and it's looking better than ever:

under down the mother and the world walks the water her,
     but dont say the probather is morning in the
     back it to the pife-started yours art
     it is the slowed to take
     from song and
     that the room and a small man.
     his sware shit in the window they she said, drever and lucked outside
     bet my began for much alone
     and the sames and now poems
     of the man came and made drisions of the corry and somebody had comples.
     I say and not there it was a chefoppectrown them, he walk the day
     bull the boy.
     perfects,
     are you write feet
     and then the world just there
     togethers? and thinking   
and this is only after the 6th epic, out of 50. The good news is it is also going a lot faster now.

Wednesday, October 5, 2016

Training Checkpoint #5

It just starting to seem a bit like some funky dialect.  We're just at about a third of the way through training.

     the birland is on shatked laten to the mother led in the hot bis
     the thilked and
     at franks and dadlen
     of short wope as other hor
     the wist rans that
     to do who got not
     of is wings and moded of the damled the watting wrote
     the tanan sorn to the altressed the atlow
     they say the boles of merens
     and the odes the tentetoting women the or pheated
     that old feal brours so louth and they.
     I shant the stage, stet pheented no
     the pirlen the theing alned me mother the toind
     and ads a stol hot felt danger and worse

At higher temperatures it's still a lot less consistent:

     I row
    and dry hot nalced I so pet
    there soind crilger lust.
     haot yey like Wadked a overylitude-awcole a phoser
    doht and nows arard it.         
      troas this want ipsrimes senrliin leumoriyrirfs gob
    comie pleblet of in k srunr basdened
    
     duanenlituects hike than other it the pans.
     he cant Henunds arastere) skatight and mofmerhoons and selr Jilz, masodor matling and
     one, I somesage to stops for
    he later of the with there worgens


Tuesday, October 4, 2016

Training Checkpoint #4

More consistent structure but still not much sense in word pairings. We're about a fifth of the way through training. 

that and  lithed dotive under tipss with rotlen
     for I the thee
     reaguns day nitispers
     we his wats
     the desetian the sturted
     the arhaswams
     and a hatading the smone on.
     jeeid
     and the seinped
     he by wold it the
     lome and
     cise, wish bonded dending
     sdree keing stonight the tutd
     terk me
     lanting to shake the coresmeds.
     where no the whute;
     we hogted her recelt she
     wusticted ad I ssenged the toith
     staded and
     were and I the of bebings
     in puntent engow and lasled the gicn
     the wound
     and she line of her,
     danned sisler on
     put duve balked;
     you there stels, turthed these, in they boind.
     is hians
     as the resjery watas. I sanns and reold and
     brank.
     I the nesd to ninled wourd steak in a bape dear stet
     old the wenking to holt.
     one the wires wover doth it and
     yang and hatch of in the most firts coog, shy
     they some mritking moncise of
     bag the oother
     starle hover
     persing can wlent yer the nitetatal thirked you erterteing under anmundy.
     it the moming and the tosish They sleak and has the bamed
     on the cast, lofky
     and jutass the thike cay hirk,
     staoded of one who
     old me how of I thilgeyer afeed on veule.
       and and wilt lanled thy wonder at
     a meon
     geing arautes,
     fol, sirdy
     bront sart
     the ineking bang sircesd poil
     encosning
     wanfongled; it lort.
     she drecolt to the tis of me sand balkis
     it waten for meress
     one the decings
     and font.
       the woltess wanted this shaging

Some of the more common words in the poetry training file are becoming obvious: she, the, I, and, he, wish, want.

Monday, October 3, 2016

Training Epoch #3

Another checkpoint milestone passed. Value loss of 1.87:

I she the way lort open
     there male,
     at in chanly, now
     world of rading world and stala and the traced as the gold
     the case and the same it is the waited to ir
     at one old sleans the youthle me
     as she sals
     and the work me in sit
     the sae
     lated the wild me
     her.
       seon want chind were the hyoe the budrare hy pewpight woet
     there lebt
     the wouse prenner detner ago mother hit and or where I hut the lock hire bacg ago
     cheeds old at a totted bown, as white tarl.
     I dealn lult in out the must wing, I stack the aind tided and she the
     met,
     ought us thas
     wanted
     when the thalk the doung the boncite and betles,
     and the raced


Still a long way to go but it's already looking better.

Training Post #2 - 10% Trained

It's a bout 1 / 10th of the way trained now.

to wonded there the asd.
 are oh got  is wowmeding of cans
     lers of rebening and enetal in to a home the jew that
     and death
     than labeing.
     it enpsrewap
     some wene tere
     the tolpy hugh
     unther cuden when on a cergusuds I rean
     on me nalde? momed their leslense gust,
     whes no were lofuden the onfokecote of she and
     to the lark, dit to vever mad.
    she wote and oe Salthere.
     we nighred the old spepes-shy.
     Chislers. the blour cusret.
     of mother
     eet and
     wote iussed?
You can see there is just a bit more clarity at this stage. There are bits of punctuation marks that it's learning and it's a little more structured.

At the lower temperature it's still just a primitive collection of words it feels more confident about:

and the borted the theat a work the thowen the were the dand the foan the and the theek the torl the tonter the and the theet the hean the want couth word of the sarled the thean bove sit and the tire the longed the that the toud bour he the hamer the forte and the datt the thenen when the thilger


Sunday, October 2, 2016

Retraining Progress

Updated the software, hardware and training file for the neural network. We'll be posting poems generated as the it trains over the next few days.  Hopefully you'll see the increase in readability.

this is after 2600 iterations with a temperature of 0.9

Q qenepchiss
     wopipoing and
     and thind are snows
     rid wish
     to sseund brotkey led
     mirle mranky
     mone, cetefen the how. I cothasfen wonge
     white.
     ceae
     bere sasly socet,sow. and hore gef ky whote ad I ssefrroir this ankocler
     in geuld at othet
     and in a lunger reeces on. he tetsen syead he kilsing and roasot annbelgal
     and a shusleces
     yor cos
     and gaige wy shangar wad roted the meike it. the teest woan Iind the nononinl,, oty geach
     and she tay uf, moderertovens't
     sits whinr a
 is wotes

on the lower end of the temperature scale, we get a sense of what words the neural network is getting confident with.

And and core the sat the and
     the and the torled the cosan the the the shele the shone on and and the the the the she the soit the the the sanler a the the the the the the and where the thes hone and the the the the the tond the the leen mone the heund the bent on the the the the the the the were the the and and the the the the the the worte the and the the the wind the the the the the the and the lorgen the there the the the wide the the the at in and the mean the lot the watting the the the conder he the bour and the net on the there the shate chart a sat cele he she the the the the cat the the and the woth her and the there the there and the the and and the the man and shoce she the the the the the the gele the the and she the the the the the sand the onsed the the wet