.. image:: https://travis-ci.org/lmjohns3/theanets.svg?branch=master .. image:: https://coveralls.io/repos/lmjohns3/theanets/badge.svg?branch=master :target: https://coveralls.io/r/lmjohns3/theanets?branch=master .. image:: http://depsy.org/api/package/pypi/theanets/badge.svg :target: http://depsy.org/package/python/theanets
THEANETSThe theanets package is a deep learning and neural network toolkit. It is
written in Python to interoperate with excellent tools like numpy and
scikit-learn, and it uses Theano_ to accelerate computations when possible
using your GPU. The package aims to provide:
The package strives to "make the easy things easy and the difficult things possible." Please try it out, and let us know what you think!
.. _Theano: http://deeplearning.net/software/theano/
Install the latest published code using pip::
pip install theanetsOr download the current source and run it from there::
git clone http://github.com/lmjohns3/theanets
cd theanets
python setup.py developNote that the contents of this README are linked to the version of theanets
that you are using. If you're looking at the README on GitHub, you'll need
the current GitHub code; if you're looking at the README for a pypi release,
you'll need the code associated with that release.
.. _README on GitHub: https://github.com/lmjohns3/theanets .. _README for a pypi release: https://pypi.python.org/pypi/theanets
Suppose you want to create a classifier and train it on some 100-dimensional data points that you've classified into 10 categories. No problem! With just a few lines you can (a) provide some data, (b) build and (c) train a model, and (d) evaluate the model:
.. code:: python
import theanets from sklearn.datasets import make_classification from sklearn.metrics import confusion_matrix
X, y = make_classification( n_samples=3000, n_features=100, n_classes=10, n_informative=10) X = X.astype('f') y = y.astype('i') cut = int(len(X) * 0.8) # training / validation split train = X[:cut], y[:cut] valid = X[cut:], y[cut:]
net = theanets.Classifier(layers=[100, 10])
net.train(train, valid, algo='sgd', learning_rate=1e-4, momentum=0.9)
for label, (X, y) in (('training:', train), ('validation:', valid)): print(label) print(confusion_matrix(y, net.predict(X)))
The model above is quite simplistic! Make it a bit more sophisticated by adding a hidden layer:
.. code:: python
net = theanets.Classifier([100, 1000, 10])
In fact, you can just as easily create 3 (or any number of) hidden layers:
.. code:: python
net = theanets.Classifier([100, 1000, 1000, 1000, 10])
By default, hidden layers use the relu transfer function. By passing a tuple instead of just an integer, you can change some of these layers to use different activations_:
.. code:: python
maxout = (1000, 'maxout:4') # maxout with 4 pieces. net = theanets.Classifier([ 100, 1000, maxout, (1000, 'tanh'), 10])
.. _activations: http://theanets.readthedocs.org/en/latest/api/activations.html
By passing a dictionary instead, you can specify even more attributes of each layer_, like how its parameters are initialized:
.. code:: python
foo = dict(name='foo', size=1000, std=1, sparsity=0.9) net = theanets.Classifier([ 100, foo, (1000, 'maxout:4'), (1000, 'tanh'), 10])
.. _layer: http://theanets.readthedocs.org/en/latest/api/layers.html
Specifying layers is the heart of building models in theanets. Read more
about this in Specifying Layers_.
.. _Specifying Layers: http://localhost:8080/guide.html#guide-creating-specifying-layers
Adding regularizers is easy, too! Just pass them to the training method. For instance, you can train up a sparse classification model with weight decay:
.. code:: python
net.train(train, valid, hidden_l1=0.001, weight_l2=0.001)
In theanets dropout is treated as a regularizer and can be set on many
layers at once:
.. code:: python
net.train(train, valid, hidden_dropout=0.5)
or just on a specific layer:
.. code:: python
net.train(train, valid, dropout={'foo:out': 0.5})
Similarly, you can add Gaussian noise to any of the layers (here, just to the input layer):
.. code:: python
net.train(train, valid, input_noise=0.3)
You can optimize your model using any of the algorithms provided by downhill
(SGD, NAG, RMSProp, ADADELTA, etc.), or additionally using a couple of
pretraining methods specific to neural networks.
.. _downhill: http://downhill.readthedocs.org/ .. _pretraining methods: http://theanets.readthedocs.org/en/latest/api/trainers.html
You can also make as many successive calls to train() as you like. Each call can include different training algorithms:
.. code:: python
net.train(train, valid, algo='rmsprop') net.train(train, valid, algo='nag')
different learning hyperparameters:
.. code:: python
net.train(train, valid, algo='rmsprop', learning_rate=0.1) net.train(train, valid, algo='rmsprop', learning_rate=0.01)
and different regularization hyperparameters:
.. code:: python
net.train(train, valid, input_noise=0.7) net.train(train, valid, input_noise=0.3)
Training models is a bit more art than science, but theanets tries to make
it easy to evaluate different training approaches. Read more about this in
Training a Model_.
.. _Training a Model: http://theanets.readthedocs.org/en/latest/guide.html#guide-training
Recurrent neural networks are becoming quite important for many sequence-based tasks in machine learning; one popular toy example for recurrent models is to generate text that's similar to some body of training text.
In these models, a recurrent classifier is set up to predict the identity of the
next character in a sequence of text, given all of the preceding characters. The
inputs to the model are the one-hot encodings of a sequence of characters from
the text, and the corresponding outputs are the class labels of the subsequent
character. The theanets code has a Text_ helper class that provides easy
encoding and decoding of text to and from integer classes; using the helper
makes the top-level code look like:
.. code:: python
import numpy as np, re, theanets
chars = re.sub(r'\s+', ' ', open('corpus.txt').read().lower()) txt = theanets.recurrent.Text(chars, min_count=10) A = 1 + len(txt.alpha) # of letter classes
net = theanets.recurrent.Classifier([A, (100, 'gru'), (1000, 'relu'), A])
seed = txt.encode(txt.text[300017:300050]) for tm, _ in net.itertrain(txt.classifier_batches(100, 32), momentum=0.9): print('{}|{} ({:.1f}%)'.format( txt.decode(seed), txt.decode(net.predict_sequence(seed, 40)), 100 * tm['acc']))
This example uses several features of theanets that make modeling neural
networks fun and interesting. The model uses a layer of Gated Recurrent Units
to capture the temporal dependencies in the data. It also uses a callable to
provide data to the model, and takes advantage of iterative training_ to
sample an output from the model after each training epoch.
.. _Text: http://theanets.readthedocs.org/en/latest/api/generated/theanets.recurrent.Text.html .. _Gated Recurrent Units: http://theanets.readthedocs.org/en/latest/api/generated/theanets.layers.recurrent.GRU.html .. _uses a callable: http://downhill.readthedocs.org/en/stable/guide.html#data-using-callables .. _iterative training: http://downhill.readthedocs.org/en/stable/guide.html#iterative-optimization
To run this example, download a text you'd like to model (e.g., Herman
Melville's Moby Dick) and save it in corpus.txt::
curl http://www.gutenberg.org/cache/epub/2701/pg2701.txt > corpus.txt
Then when you run the script, the output might look something like this (abbreviated to show patterns)::
used for light, but only as an oi|pr vgti ki nliiariiets-a, o t.;to niy , (16.6%) used for light, but only as an oi|s bafsvim-te i"eg nadg tiaraiatlrekls tv (20.2%) used for light, but only as an oi|vetr uob bsyeatit is-ad. agtat girirole, (28.5%) used for light, but only as an oi|siy thinle wonl'th, in the begme sr"hey (29.9%) used for light, but only as an oi|nr. bonthe the tuout honils ohe thib th (30.5%) used for light, but only as an oi|kg that mand sons an, of,rtopit bale thu (31.0%) used for light, but only as an oi|nsm blasc yan, ang theate thor wille han (32.1%) used for light, but only as an oi|b thea mevind, int amat ars sif istuad p (33.3%) used for light, but only as an oi|msenge bie therale hing, aik asmeatked s (34.1%) used for light, but only as an oi|ge," rrermondy ghe e comasnig that urle (35.5%) used for light, but only as an oi|s or thartich comase surt thant seaiceng (36.1%) used for light, but only as an oi|s lot fircennor, unding dald bots trre i (37.1%) used for light, but only as an oi|st onderass noptand. "peles, suiondes is (38.2%) used for light, but only as an oi|gnith. s. lited, anca! stobbease so las, (39.3%) used for light, but only as an oi|chics fleet dong berieribus armor has or (40.1%) used for light, but only as an oi|cs and quirbout detom tis glome dold pco (41.1%) used for light, but only as an oi|nht shome wand, the your at movernife lo (42.0%) used for light, but only as an oi|r a reald hind the, with of the from sti (43.0%) used for light, but only as an oi|t beftect. how shapellatgen the fortower (44.0%) used for light, but only as an oi|rtucated fanns dountetter from fom to wi (45.2%) used for light, but only as an oi|r the sea priised tay queequings hearhou (46.8%) used for light, but only as an oi|ld, wode, i long ben! but the gentived. (48.0%) used for light, but only as an oi|r wide-no nate was him. "a king to had o (49.1%) used for light, but only as an oi|l erol min't defositanable paring our. 4 (50.0%) used for light, but only as an oi|l the motion ahab, too, and relay in aha (51.0%) used for light, but only as an oi|n dago, and contantly used the coil; but (52.3%) used for light, but only as an oi|l starbuckably happoss of the fullies ti (52.4%) used for light, but only as an oi|led-bubble most disinuan into the mate-- (53.3%) used for light, but only as an oi|len. ye?' 'tis though moby starbuck, and (53.6%) used for light, but only as an oi|l, and the pequodeers. but was all this: (53.9%) used for light, but only as an oi|ling his first repore to the pequod, sym (54.4%) used for light, but only as an oi|led escried; we they like potants--old s (54.3%) used for light, but only as an oi|l-ginqueg! i save started her supplain h (54.3%) used for light, but only as an oi|l is, the captain all this mildly bounde (54.9%)
Here, the seed text is shown left of the pipe character, and the randomly sampled sequence follows. In parantheses are the per-character accuracy values on the training set while training the model. The pattern of learning proceeds from almost-random character generation, to producing groups of letters separated by spaces, to generating words that seem like they might belong in Moby Dick, things like "captain," "ahab, too," and "constantly used the coil."
Much amusement can be derived from a temporal model extending itself forward in this way. After all, how else would we ever think of "Pequodeers," "Starbuckably," or "Ginqueg"?!
Source: https://github.com/lmjohns3/theanets
Documentation: http://theanets.readthedocs.org
Mailing list: https://groups.google.com/forum/#!forum/theanets