PyTorch Poetry Language Model.
Trained on approx 600,000 lines of poetry
jhave@jhave-Ubuntu:~/Documents/Github/pytorch-poetry-generation/word_language_model$ python generate_2017-INFINITE-1M.py –cuda –checkpoint=’/home/jhave/Documents/Github/pytorch-poetry-generation/word_language_model/models/2017-06-17T09-22-17/model-LSTM-emsize-1860-nhid_1860-nlayers_2-batch_size_20-epoch_30-loss_6.00-ppl_405.43.pt’
Embedding size: 1860
Hidden Layers: 1860
Batch size: 20
jhave@jhave-Ubuntu:~/Documents/Github/pytorch-poetry-generation/word_language_model$ python generate_2017-INFINITE-1M.py –cuda –checkpoint=’/home/jhave/Documents/Github/pytorch-poetry-generation/word_language_model/models/2017-06-17T09-22-17/model-LSTM-emsize-1860-nhid_1860-nlayers_2-batch_size_20-epoch_31-loss_6.00-ppl_405.39.pt’
Embedding size: 1860
Hidden Layers: 1860
Batch size: 20
The train lost, “Aye man!
O old beggar, O perfect friend;
The bath-tub before the Bo’s’n resentments; pissing
rimed metaphors in the white pincers
scratching and whiten each of a clarity
in the sky the sacred hoof of eastward,
arc of the pestle through sobered the cliffs to the own world.
Personally I like the coherence of Pytorch, it’s capacity to hold the disembodied recalcitrant veil of absurdity over a somewht stoic normative syntactical model.
Jumble, Rub Up The Him-Whose-Penis-Stretches-Down-To-His-Knees. The slow-wheeling white Thing withheld in the light of the Whitman? The roof kisses the wounds of blues and yellow species. Far and cold in the soft cornfields bending to the gravel, Or showing diapers in disclosure, Atlantic, Raymond Protract the serried ofercomon, â€” the throats "I've used to make been sustene, Fanny, the inner man clutched to the keep; Who meant me to sing one step at graves.
weather, Gold apply thy unndmight hour. And neither would endure Meet excel understood) I once had declared clay. Be lines once my God
accepting tall any flowers, forever with one question
of boots, neural, dead forgotten the glass
of cloud, start and more, who studied legends
and wanted to ascend
Every inch alone you and this desire
tulips of sounds
watching the witness
On the intensity lolling it.
Summer up warmishment
The girls crack our hearts
she set and quickens, swarms at the edge.
where they could wake? It begins
how much design, anthounces are taught her.
Illusion, mimicry a chalk.
when you’re lonely, black large in the calico,
where the bachelo forking genes
might in dusty confidently,
ignore and suck with the main grove,
the dream in the darkness, found hear
if these wobbles, silver for the man.
Deaf-soot he’s an edging of ships
a border that is the court.
who wants shapelesse lava
So long as nothing moved oversized no mountains of an eternity
I fed him into oned
There is to say something a fog and mask at writing
minild, the moon and cair of his screens
It is there learning, loving down, screeching.
mystery, painted Spring.
as mid-afternoon, fetlocks
uncurledding cheaping full of pale
eternal, grabs us. Flowers try
migrating every idea
and whispered at the
looking at her skin, burst
from her will.
The opposite of scarcity is not abundance: saturation inevitably damages, or perhaps it just corrodes or blunts the impassioned pavlov puppy, delivering a dent of tiny deliberate delirious wonder.
Yet technical momentum now propels and impels continuance: how can the incoherence be tamed? Set some new hyperparameters and let the wvnt train for 26 hours.
Over this weekend, I’ve churned out about 100,000 lines. Generating reckless amounts of incoherent poetry threatens more than perspective or contemplation, it totters sanity on the whim of a machine. Teeming bacteria, every epiphany a whiff of redundancy.
$ python train_2017_py3p5_150k_low6_sample4096_SAVE-STARTS_100k.py --wavenet_params=wavenet_params_ORIG_dilations2048_skipChannels8192_qc2048_dc32.json --data_dir=data/2017 Using default logdir: ./logdir/train/2017-06-01T08-38-45 _______________________________________________________________________________________________ dilations: 2048 filter_width: 2 residual_channels: 32 dilation_channels: 32 skip_channels: 8192 quantization_channels: 2048
and faithful wound To fruit white, the dread One by one another, Image saved-- Ay of the visit. What pursued my heart to brink. Such the curse of hopes fraught memory;
Caught a new light bulb, All the heart is grown.
TXTs generated in SLOW MODE
There’s a way of calculating the matrices that taxes the strength of evn a magnificient GPU, making production crawl, and the computer difficult to use. Each of the following txt files (4444 letters in each) took about 40-60 minutes to generate on an Nvidia Maxwell TitanX using cuda8 on Ubuntu 16.4
Txts generated slow seem somehow thicker, as if issued from a more calibrated mentation, yet at the same time it’s math scat, glitch flow. Glisses from disintegrating encyclopedias.
Here are some samples:
I found myself within us wonder. You purchase as ease with water over events, because the straightforward that I miximally, she Don't sports commentation with its ruffled story
So the abrasive brash gutter voice of the neural net seemed maybe due to lack of longterm exposure, wider horizons, deeper reading into the glands of embodied organisms, so I set the hyperparameters higher, waited 26 hours and watched the ubuntu HD fill up with models to the point where the OS crashed on reboot and i found myself entering a mysterious cmd line universe called grub… thus to say apprenticing a digital poet is not without perils.
The title of this post is the title of the first poem generated by a re-installed Wavenet for Poetry Generation (still using 0.1 Tensorflow, but now on Python 3.5), and working on an expanded corpus (using approx 600,000 lines of poetry) the first model was Model: 7365 | Loss: 0.641 | Trained on: 2017-05-27T14-19-11 (full txt here).
Wavenet is a rowdy poet, jolting neologisms, careening rhythms, petulant insolence, even the occasional glaring politically-incorrect genetic smut. Tangents codified into contingent unstable incoherence.
Compared to Pytorch, which aspires to a refined smooth reservoir of cadenced candy, Wavenet is a drunken street brawler: rude, slurring blursh meru crosm nanotagonisms across rumpled insovite starpets.
Pytorch is Wallace Stevens; Wavenet is Bukowski (if he’d been born a mathematician neologist).
Here’s a random poem:
Model: 118740 | Loss: 0.641 | 2017-05-28T00-35-33
Eyes calm, nor or something cases.
from a wall coat hardware as it times, a fuckermarket
in my meat by the heart, earth signs: a pupil, breaths &
stops children, pretended. But they were.
Case study: Folder 2017-05-28T12-16-50 contains 171 models (each saved because their loss was under the 0.7 threshold). But what does loss really mean? In principle it is a measurement of the gap between the generated text and the validation text (how close is it?); yet however many different schemas proliferate, loss (like pain) cannot be measured by instrumentality.
Here’s another random poem:
Model: 93286 | Loss: 0.355 | 2017-05-28T12-16-50
would destroying through the horizon. The poor
Sequel creation rose sky.
So we do not how you bastard, grew,
there is no populously, despite bet.
Trees me that he went back
on tune parts.
I will set
a girl of sunsets in the glass,
and no one even on the floral came
I’m slowly learning the hard way to wrap each install using VirtualEnvironments. Without that as the upgrades happen, code splinters and breaks, leaking a fine luminous goo of errors.
The current virtual environment was created using
$ sudo apt-get install python3-pip python3-dev python-virtualenv $ virtualenv -p python3.5 ~/tf0p1-py3.5-wvnt
After that, followed the instructions,
$ TF_BINARY_URL=https//storage.googleapis.com/tensorflow/mac/gpu/tensorflow-0.10.0-py3-none-any.whl $ pip3 install --upgrade $TF_BINARY_URL
then got snarled into a long terrible struggle with error messages messing up the output, resolved by inserting,
os.environ['TF_CPP_MIN_LOG_LEVEL']='2' # into generate_Poems_2017-wsuppressed.py
And to generate on Ubuntu, using the lovely Nvidia Titan X GPU so generously donated by Nvidia under their academic grant program:
$ cd Documents/Github/Wavenet-for-Poem-Generation/ $ source ~/tf0p1-py3.5-wvnt/bin/activate (tf0p1-py3.5-wvnt)$ python train_2017_py3p5.py --data_dir=data/2017 --wavenet_params=wavenet_params_ORIG_dilations1024_skipChannels4096_qc1024_dc32.json
tf0p1-py3.5-wvnt_jhave-Ubuntu_Screencast 2017-05-28 11:31:40_TrainedOn_2017-05-28T00-35-33 tf0p1-py3.5-wvnt_jhave-Ubuntu_2017-05-28 09:18:00_TRAINED_2017-05-28T00-35-33 tf0p1-py3.5-wvnt_jhave-Ubuntu_Screencast-2017-05-27 23:50:14_basedon_2017-05-27T14-19-11 tf0p1-py3.5-wvnt_jhave-Ubuntu_Screencast 2017-05-28 22:36:35_TrainedOn_2017-05-28T12-16-50.txt
Mead recipe: dilute honey with water, stir twice a day, wait.
Fermentation begins after 24-48 hours. After a week, the fermented honey-wine (mead) can be enjoyed green, low in alcohol yet lively with essence. Or you can let it continue.
Generative-poetry recipe: text-corpus analysed by neural net, wait.
After each reading of the corpus (a.k.a. ‘training epoch’), a neural net can produce/save a model. Think of the model as an ecosystem produced by fermentation, an idea, a bubbling contingency, a knot of flavours, a succulent reservoir, a tangy substrate that represents the contours, topology or intricacies of the source corpus. Early models ( a model is usually produced after the algorithm reads the entire corpus; remember, each reading of the corpus is called an epoch; after 4-7 epochs the models are still young) may be obscure but are often buoyant with energy (like kids mimicking modern dance).
Here’s an example from a model that is only 2 epochs old:
Daft and bagel, of sycamore.
“Now thou so sex–sheer if I see Conquerours.
So he is beneath the lake, where they pass together,
Amid the twister that squeeze-box fire:
Here’s an example from a model that is 6 epochs old:
Yellow cornbread and yellow gems
and all spring to eat.
Owls of the sun: the oldest worm of light
robed in the advances of a spark.
Later models (after 20 epochs), held in vats (epoch after epoch), exhibit more refined microbial (i.e. algorithm) calibrations:
Stale nose of empty lair
The bright lush walls are fresh and lively and weeping,
Whirling into the night as I stand
unicorn raging on the serenade
Of floors of sunset after evil-spirits,
Eventually fermentation processes halt. In the case of mead, this may occur after a month; with neural nets, a process called simulated annealing intentionally decreases the learning rate every iteration; so the system begins by exploring large features then focuses on details. Eventually the learning rate diminishes to zero. Learning (fermentation) stops.
2017-04-30 09:49:10_EPOCH-6 2017-04-30 13:27:15_EPOCH-16 2017-04-30 23:03:42_EPOCH-4 2017-05-03 09:41:35_2017-05-02T06-19-17_EPOCH22 2017-05-03 12:54:57_2017-05-02T06-19-17_EPOCH-6 2017-05-12 12:24:54_2017-05-02T06-19-17_EPOCH22
2017-05-13 21:24:06_2017-05-02T06-19-17_-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_7-loss_6.02-ppl_412.27 2017-05-14 23:51:14_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_8-loss_6.03-ppl_417.49 2017-05-15 11:00:03_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_3-loss_6.09-ppl_439.57 2017-05-15 11:28:24_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_4-loss_6.04-ppl_421.71 2017-05-15 11:53:40_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_5-loss_6.02-ppl_413.64 2017-05-15 21:08:56_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_9-loss_6.02-ppl_409.87 2017-05-15 21:48:55_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_10-loss_6.03-ppl_416.36 2017-05-15 22:21:11_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_11-loss_6.02-ppl_413.61 2017-05-15 22:48:53_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_12-loss_6.03-ppl_414.69 2017-05-15 23:24:25_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_13-loss_6.03-ppl_417.08 2017-05-15 23:53:21_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_14-loss_6.03-ppl_416.68 pytorch-poet_2017-05-02T06-19-17_models-6-7-8-9-10-11-12-13-14 2017-05-16 09:26:18_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_15-loss_6.03-ppl_416.602017-05-16 10:08:15_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_16-loss_6.03-ppl_414.912017-05-16 10:45:58_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_17-loss_6.03-ppl_415.35 2017-05-18 23:48:36_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_18-loss_6.03-ppl_414.102017-05-20 10:35:17_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_19-loss_6.03-ppl_413.682017-05-18 22:10:36_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_17-loss_6.03-ppl_415.35 2017-05-20 22:01:22_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_20-loss_6.02-ppl_413.082017-05-21 12:58:03_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_21-loss_6.02-ppl_413.072017-05-21 14:38:29_model-LSTM-emsize-1500-nhid_1500-nlayers_2-batch_size_20-epoch_22-loss_6.02-ppl_412.68