Skip to main content

Practical Machine Learning for NLP

Project description

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.

🔮 Version 6.10 out now! Read the release notes here.

Build Status Appveyor Build Status Test Coverage Current Release Version pypi Version conda Version Thinc on Gitter Follow us on Twitter

What’s where (as of v6.9.0)

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:

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)
        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.


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
python -c "import thinc.neural.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
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
cd thinc

virtualenv .env
source .env/bin/activate

pip install -r requirements.txt
python build_ext --inplace
py.test thinc/

And then run the examples as follows:

python examples/
python examples/
python examples/


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 models f(x) and g(x) into a single model computing g(f(x)).
  • clone(model1, int): Create n copies 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 = (
        >> flatten
        >> memoize(
            | (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
  • describe.on_data(): Allows callbacks to be specified, which will be called on Model.begin_training().

🛠 Changelog

Version Date Description
v6.10.1 2017-11-15 Fix GPU install and minor memory leak
v6.10.0 2017-10-28 CPU efficiency improvements, refactoring
v6.9.0 2017-10-03 Reorganize layers, bug fix to Layer Normalization
v6.8.2 2017-09-26 Fix packaging of gpu_ops
v6.8.1 2017-08-23 Fix Windows support
v6.8.0 2017-07-25 SELU layer, attention, improved GPU/CPU compatibility
v6.7.3 2017-06-05 Fix convolution on GPU
v6.7.2 2017-06-02 Bug fixes to serialization
v6.7.1 2017-06-02 Improve serialization
v6.7.0 2017-06-01 Fixes to serialization, hash embeddings and flatten ops
v6.6.0 2017-05-14 Improved GPU usage and examples
v6.5.2 2017-03-20 n/a
v6.5.1 2017-03-20 Improved linear class and Windows fix
v6.5.0 2017-03-11 Supervised similarity, fancier embedding and improvements to linear model
v6.4.0 2017-02-15 n/a
v6.3.0 2017-01-25 Efficiency improvements, argument checking and error messaging
v6.2.0 2017-01-15 Improve API and introduce overloaded operators
v6.1.3 2017-01-10 More neural network functions and training continuation
v6.1.3 2017-01-09 n/a
v6.1.2 2017-01-09 n/a
v6.1.1 2017-01-09 n/a
v6.1.0 2017-01-09 n/a
v6.0.0 2016-12-31 Add thinc.neural for NLP-oriented deep learning

Release history Release notifications | RSS feed

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for thinc, version 6.12.1
Filename, size File type Python version Upload date Hashes
Filename, size thinc-6.12.1-cp27-cp27m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (2.7 MB) File type Wheel Python version cp27 Upload date Hashes View
Filename, size thinc-6.12.1-cp27-cp27m-manylinux1_x86_64.whl (1.9 MB) File type Wheel Python version cp27 Upload date Hashes View
Filename, size thinc-6.12.1-cp35-cp35m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (2.6 MB) File type Wheel Python version cp35 Upload date Hashes View
Filename, size thinc-6.12.1-cp35-cp35m-manylinux1_x86_64.whl (1.9 MB) File type Wheel Python version cp35 Upload date Hashes View
Filename, size thinc-6.12.1-cp35-cp35m-win_amd64.whl (1.8 MB) File type Wheel Python version cp35 Upload date Hashes View
Filename, size thinc-6.12.1-cp36-cp36m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (2.7 MB) File type Wheel Python version cp36 Upload date Hashes View
Filename, size thinc-6.12.1-cp36-cp36m-manylinux1_x86_64.whl (1.9 MB) File type Wheel Python version cp36 Upload date Hashes View
Filename, size thinc-6.12.1-cp36-cp36m-win_amd64.whl (1.8 MB) File type Wheel Python version cp36 Upload date Hashes View
Filename, size thinc-6.12.1-cp37-cp37m-macosx_10_6_intel.macosx_10_9_intel.macosx_10_9_x86_64.macosx_10_10_intel.macosx_10_10_x86_64.whl (2.6 MB) File type Wheel Python version cp37 Upload date Hashes View
Filename, size thinc-6.12.1-cp37-cp37m-manylinux1_x86_64.whl (1.9 MB) File type Wheel Python version cp37 Upload date Hashes View
Filename, size thinc-6.12.1-cp37-cp37m-win_amd64.whl (1.8 MB) File type Wheel Python version cp37 Upload date Hashes View
Filename, size thinc-6.12.1.tar.gz (1.2 MB) File type Source Python version None Upload date Hashes View

Supported by

AWS AWS Cloud computing Datadog Datadog Monitoring Facebook / Instagram Facebook / Instagram PSF Sponsor Fastly Fastly CDN Google Google Object Storage and Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Salesforce Salesforce PSF Sponsor Sentry Sentry Error logging StatusPage StatusPage Status page