Category: data art

PyTorch is an early release beta software (developed by a consortium led by Facebook and NIVIDIA), a “deep learning software that puts Python first.”

So since I luckily received an NVIDIA GTX TitanX (Maxwell) before leaving Hong Kong under the generous NVIDIA academic GPU Grant program, and having last week finally bought a custom-build to house it, and 2 days ago finally got Ubuntu installed with CUDA and CUDNN drivers, and having found that the Tensorflow 0.11 version no longer runs under Python 3.6 Anaconda, I decided to give a PyTorch example a try, specifically Word-level language modeling RNN

This example trains a multi-layer RNN (Elman, GRU, or LSTM) on a language modeling task…The trained model can then be used by the generate script to generate new text.

And after only an hour of training on an 11k poem corpus, using the default settings, the results announced “End of training | test loss  5.99 | test ppl   398.41” — Which means that the loss is bad and perplexity is now at the seemingly terrible level of 398….

Then I ran the generate script and the 1000 word text below got generated in less than 30 seconds. I find it stunning. If this is what PyTorch is capable of with a tiny corpus, default settings and a minimal run, language generation is entering a renaissance.  Ok, so it’s veering toward the incomprehensible and has little lived evocative phenomenological resonance, but its grasp on idiomatic cadence is creepily accurate. It’s as if it absorbed several semesters of graduate seminars on romantic and post-modern verse:

the embankment
and your face sad like a nest, grew sorry
when your cold work made of snow
broken
and left a thousand magnifies.

a little cold, you plant but hold it
and seems
the slight arts? face, and ends
with such prayer as the fingers do,
this reorganizing contest is how
to be murdered
throwing it
into the arteries obscurity goes disc whispering whole
affairs, now your instinct
does a case,
defense. on her eye, you do not know that every homelands
is didn’t at the
risk very psychiatrists, just under bay.

by the living of life’s melancholy grate.
i have found a
wild orange in eden, eight hazy years guzzles
her neck at the grave turn into every mythological orbit of
distances,
person’s there–see then are we told what we understand
won’t take the slightest danger
or the
size of what it means to take up if you can,
tongue. only your eye exultant whitens again will
happen.
i think that the four-oared clouded of one stick in flowerpot
is part of an antique little
register on a hiatus
till i try for you.
i wash up the door my knee will be
high.
if i refuse a limits as i can lift my hand rubicon.

i can see her
above the stove tide
hip. orange as a breaking sty.

Continue reading “Testing PyTorch on Poems (Preliminary Results)”

For the past week, I’ve been running a port of the Wavenet algorithm to generate poems. A reasonable training result emerges in about 24 hours, — a trained model that can generate immense amounts of text relatively quickly. On a laptop. (Code: github). By reasonable I mean the poems do not have any real sense, no sentient self, no coherent narrative, nor epic structure. But they do have cadence, they do not repeat, new words are plausible, and they have adopted a scattered open line style characteristic of the late twentieth century corpus on which they were trained. Much more lucid than Schwitters’ Ursonate, output is reminiscent of Beckett’s Not I : ranting incandescent perpetual voice.

Results

Remember, these are evolutionary amoebas, toddlers just learning to babble. The amazing thing is that without being given any syntax rules, they are speaking, generating a kind of prototypical glossolalia poem, character by character. Note: models are like wines, idiosyncratic reservoirs, the output of each has a distinct taste, — some have mastered open lines, others mutter densely, many mangle words to make neologisms  — each has obsessions. The Wavenet algorithm is analogous to a winery: its processes ensure that all of the models are similar. Tensorflow is the local region; recursive neural nets form the ecosystem. The corpus is the grapes.

Intriguing vintages-models :

Dense intricate Model 33380 — trained with 1024 skip channels and dilation to 1024 (read a txt sample)

the mouth’s fruit

tiny from carrying

a generative cup

Loose uncalibrated Model 13483 with loss = 0.456, (1.436 sec/step) trained on 2016-10-15T20-46-39 with 2048 skip channels and dilation to 256 (read a txt sample)

 at night, say, that direction.

          sleeps

      now. so you hear we are shaking

          from the woods

Full results (raw output, unedited txt files from the week of Oct 10-16th 2016) here.

it’s there we brail,

  beautiful full

left to wish our autumn was floor

Code

Python source code + a few trained models, corpus and some sample txt: on github which will be updated with new samples and code as it emerges.

 Details

The Model number refers to how many steps it trained for. Skip channels weave material from different contexts. On this corpus, larger skip channels produce more coherent output. Dilations refer to the size of the tensors of the encoder-decoder: eg. [ 1, 2, 4, 8, 16, 32, 64, 128, 256, etc… ] Higher values up to 1024 seem to be of benefit, but take longer to train. Loss is the mathematical calculation of the distance between the goal and the model; it is a measure of how tightly the model fits the topological shape of the corpus; as models are trained, they are supposed to learn to minimize loss; low loss is supposed to be good. For artistic purposes this is questionable (I describe why in see next section). For best results, in general, on this corpus: 10k to 50k steps, 1024 dilations, a skip channel of 512 or more, and (most crucial) loss less than 0.6.

Learned

Loss is not everything. An early iteration model with low loss will generate cruft with immense spelling errors. Thousands of runs later, a model with the same loss value will usually produce more sophisticated variations, less errors. So there is more going on inside the system than is captured by the simple metric of loss optimization. Moreover if the system is about to  undergo a catastrophic blowout of loss values, during which the loss ceases to descend toward the gradient and exponentially oscillates (this occasionally occurs after approx 60k steps). Generated text from poems just before that (with good loss values below 1.0 or even excellent loss values below 0.6) will produce some ok stuff interspersed with long periods of nonsense or —— repeated **** symbols. These repetitive stretches are symptoms of the imminent collapse. So loss is not everything. Nonsense can be a muse. Mutating small elements, editing, flowing, falling across the suggestive force of words in raw tumult provides a viable medium for finding voice. Continue reading “Wavenet for Poem Generation: preliminary results”