Thinc: Practical Machine Learning for NLP in Python
Thinc is the machine learning library powering spaCy.
It features a battle-tested linear model designed for large sparse learning
problems, and a flexible neural network model under development for
spaCy v2.0.
Thinc is a practical toolkit for implementing models that follow the
"Embed, encode, attend, predict"
architecture. It's designed to be easy to install, efficient for CPU usage and
optimised for NLP and deep learning with text – in particular, hierarchically
structured input and variable-length sequences.
? Read the release notes here.
What's where (as of v7.0.0), Module, Description, ----------------------, -------------------------------------------------------------------, thinc.v2v.Model, Base class., thinc.v2v, Layers transforming vectors to vectors., thinc.i2v, Layers embedding IDs to vectors., thinc.t2v, Layers pooling tensors to vectors., thinc.t2t, Layers transforming tensors to tensors (e.g. CNN, LSTM)., thinc.api, Higher-order functions, for building networks. Will be renamed., thinc.extra, Datasets and utilities., thinc.neural.ops, Container classes for mathematical operations. Will be reorganized., thinc.linear.avgtron, Legacy efficient Averaged Perceptron implementation., ## Development status
Thinc's deep learning functionality is still under active development: APIs are
unstable, and we're not yet ready to provide usage support. However, if you're
already quite familiar with neural networks, there's a lot here you might find
interesting. Thinc's conceptual model is quite different from TensorFlow's.
Thinc also implements some novel features, such as a small DSL for concisely
wiring up models, embedding tables that support pre-computation and the
hashing trick, dynamic batch sizes, a concatenation-based approach to
variable-length sequences, and support for model averaging for the
Adam solver (which performs very well).
No computational graph – just higher order functions
The central problem for a neural network implementation is this: during the
forward pass, you compute results that will later be useful during the backward
pass. How do you keep track of this arbitrary state, while making sure that
layers can be cleanly composed?
Most libraries solve this problem by having you declare the forward
computations, which are then compiled into a graph somewhere behind the scenes.
Thinc doesn't have a "computational graph". Instead, we just use the stack,
because we put the state from the forward pass into callbacks.
All nodes in the network have a simple signature:
f(inputs) -> {outputs, f(d_outputs)->d_inputs}
To make this less abstract, here's a ReLu activation, following this signature:
def relu(inputs):
mask = inputs > 0
def backprop_relu(d_outputs, optimizer):
return d_outputs * mask
return inputs * mask, backprop_relu
When you call the relu function, you get back an output variable, and a
callback. This lets you calculate a gradient using the output, and then pass it
into the callback to perform the backward pass.
This signature makes it easy to build a complex network out of smaller pieces,
using arbitrary higher-order functions you can write yourself. To make this
clearer, we need a function for a weights layer. Usually this will be
implemented as a class — but let's continue using closures, to keep things
concise, and to keep the simplicity of the interface explicit.
The main complication for the weights layer is that we now have a side-effect
to manage: we would like to update the weights. There are a few ways to handle
this. In Thinc we currently pass a callable into the backward pass. (I'm not
convinced this is best.)
import numpy
def create_linear_layer(n_out, n_in):
W = numpy.zeros((n_out, n_in))
b = numpy.zeros((n_out, 1))
def forward(X):
Y = W @ X + b
def backward(dY, optimizer):
dX = W.T @ dY
dW = numpy.einsum('ik,jk->ij', dY, X)
db = dY.sum(axis=0)
optimizer(W, dW)
optimizer(b, db)
return dX
return Y, backward
return forward
If we call Wb = create_linear_layer(5, 4), the variable Wb will be the
forward() function, implemented inside the body of create_linear_layer().
The Wb instance will have access to the W and b variable defined in its
outer scope. If we invoke create_linear_layer() again, we get a new instance,
with its own internal state.
The Wb instance and the relu function have exactly the same signature. This
makes it easy to write higher order functions to compose them. The most obvious
thing to do is chain them together:
def chain(*layers):
def forward(X):
backprops = []
Y = X
for layer in layers:
Y, backprop = layer(Y)
backprops.append(backprop)
def backward(dY, optimizer):
for backprop in reversed(backprops):
dY = backprop(dY, optimizer)
return dY
return Y, backward
return forward
We could now chain our linear layer together with the relu activation, to
create a simple feed-forward network:
Wb1 = create_linear_layer(10, 5)
Wb2 = create_linear_layer(3, 10)
model = chain(Wb1, relu, Wb2)
X = numpy.random.uniform(size=(5, 4))
y, bp_y = model(X)
dY = y - truth
dX = bp_y(dY, optimizer)
This conceptual model makes Thinc very flexible. The trade-off is that Thinc is
less convenient and efficient at workloads that fit exactly into what
TensorFlow etc. are designed for. If your graph
really is static, and your inputs are homogenous in size and shape,
Keras will likely be faster and simpler. But if you want
to pass normal Python objects through your network, or handle sequences and
recursions of arbitrary length or complexity, you might find Thinc's design a
better fit for your problem.
Quickstart
Thinc should install cleanly with both pip
and conda, for
Pythons 2.7+ and 3.5+, on Linux, macOS / OSX and Windows. Its
only system dependency is a compiler tool-chain (e.g. build-essential) and
the Python development headers (e.g. python-dev).
pip install thinc
For GPU support, we're grateful to use the work of Chainer's cupy module,
which provides a numpy-compatible interface for GPU arrays. However, installing
Chainer when no GPU is available currently causes an error. We therefore do not
list Chainer as an explicit dependency — so building Thinc for GPU requires
some extra steps:
export CUDA_HOME=/usr/local/cuda-8.0 # Or wherever your CUDA is
export PATH=$PATH:$CUDA_HOME/bin
pip install chainer
python -c "import cupy; assert cupy" # Check it installed
pip install thinc_gpu_ops thinc # Or `thinc[cuda]`
python -c "import thinc_gpu_ops" # Check the GPU ops were built
The rest of this section describes how to build Thinc from source. If you have
Fabric installed, you can use the shortcut:
git clone https://github.com/explosion/thinc
cd thinc
fab clean env make test
You can then run the examples as follows:
fab eg.mnist
fab eg.basic_tagger
fab eg.cnn_tagger
Otherwise, you can build and test explicitly with:
git clone https://github.com/explosion/thinc
cd thinc
virtualenv .env
source .env/bin/activate
pip install -r requirements.txt
python setup.py build_ext --inplace
py.test thinc/
And then run the examples as follows:
python examples/mnist.py
python examples/basic_tagger.py
python examples/cnn_tagger.py
Usage
The Neural Network API is still subject to change, even within minor versions.
You can get a feel for the current API by checking out the examples. Here are
a few quick highlights.
1. Shape inference
Models can be created with some dimensions unspecified. Missing dimensions are
inferred when pre-trained weights are loaded or when training begins. This
eliminates a common source of programmer error:
# Invalid network — shape mismatch
model = chain(ReLu(512, 748), ReLu(512, 784), Softmax(10))
# Leave the dimensions unspecified, and you can't be wrong.
model = chain(ReLu(512), ReLu(512), Softmax())
2. Operator overloading
The Model.define_operators() classmethod allows you to bind arbitrary
binary functions to Python operators, for use in any Model instance. The
method can (and should) be used as a context-manager, so that the overloading
is limited to the immediate block. This allows concise and expressive model
definition:
with Model.define_operators({'>>': chain}):
model = ReLu(512) >> ReLu(512) >> Softmax()
The overloading is cleaned up at the end of the block. A fairly arbitrary zoo
of functions are currently implemented. Some of the most useful:
chain(model1, model2): Compose two modelsf(x)andg(x)into a single model computingg(f(x)).clone(model1, int): Createncopies of a model, each with distinct weights, and chain them together.concatenate(model1, model2): Given two models with output dimensions(n,)and(m,), construct a model with output dimensions(m+n,).add(model1, model2):add(f(x), g(x)) = f(x)+g(x)make_tuple(model1, model2): Construct tuples of the outputs of two models, at the batch level. The backward pass expects to receive a tuple of gradients, which are routed through the appropriate model, and summed.
Putting these things together, here's the sort of tagging model that Thinc is
designed to make easy.
with Model.define_operators({'>>': chain, '**': clone, ', ': concatenate}):
model = (
add_eol_markers('EOL')
>> flatten
>> memoize(
CharLSTM(char_width), (normalize >> str2int >> Embed(word_width)))
>> ExtractWindow(nW=2)
>> BatchNorm(ReLu(hidden_width)) ** 3
>> Softmax()
)
Not all of these pieces are implemented yet, but hopefully this shows where
we're going. The memoize function will be particularly important: in any
batch of text, the common words will be very common. It's therefore important
to evaluate models such as the CharLSTM once per word type per minibatch,
rather than once per token.
3. Callback-based backpropagation
Most neural network libraries use a computational graph abstraction. This takes
the execution away from you, so that gradients can be computed automatically.
Thinc follows a style more like the autograd library, but with larger
operations. Usage is as follows:
def explicit_sgd_update(X, y):
sgd = lambda weights, gradient: weights - gradient * 0.001
yh, finish_update = model.begin_update(X, drop=0.2)
finish_update(y-yh, sgd)
Separating the backpropagation into three parts like this has many advantages.
The interface to all models is completely uniform — there is no distinction
between the top-level model you use as a predictor and the internal models for
the layers. We also make concurrency simple, by making the begin_update()
step a pure function, and separating the accumulation of the gradient from the
action of the optimizer.
4. Class annotations
To keep the class hierarchy shallow, Thinc uses class decorators to reuse code
for layer definitions. Specifically, the following decorators are available:
describe.attributes(): Allows attributes to be specified by keyword argument. Used especially for dimensions and parameters.describe.on_init(): Allows callbacks to be specified, which will be called at the end of the__init__.py.describe.on_data(): Allows callbacks to be specified, which will be called onModel.begin_training().
