# Layers

This page describes functions for defining your model. Each layer is implemented
in its own module in
`thinc.layers`

and can be imported from `thinc.api`

. Most layer files define two public
functions: a **creation function** that returns a `Model`

instance, and a **forward function** that performs the computation.

Weights layers | Layers that use an internal weights matrix for their computations. |

Reduction operations | Layers that perform rank reductions, e.g. pooling from word to sentence vectors. |

Combinators | Layers that combine two or more existing layers. |

Data type transfers | Layers that transform data to different types. |

Wrappers | Wrapper layers for other libraries like PyTorch and TensorFlow. |

## Weights layers

### CauchySimilarity function

Compare input vectors according to the Cauchy similarity function proposed by
Chen (2013).
Primarily used within `siamese`

neural networks.

Argument | Type | Description |
---|---|---|

`nI` | `Optional[int]` | The size of the input vectors. |

RETURNS | `Model[Tuple[Floats2d, Floats2d], Floats1d]` | The created similarity layer. |

View on GitHub## thinc/layers/cauchysimilarity.py

`Can't fetch code example from GitHub :( Please use the link above to view the example. If you've come across a broken link, we always appreciate a pull request to the repository, or a report on the issue tracker. Thanks!`

### Dish function

A dense layer with the Dish activation function. Dish or “Daniël’s Swish-like
activation” is an activation function with a non-monotinic shape similar to
GELU, Swish and Mish. However, Dish does not rely on
elementary functions like `exp`

or `erf`

, making it much
faster to compute
in most cases.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `he_normal_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/dish.py

`Can't fetch code example from GitHub :( Please use the link above to view the example. If you've come across a broken link, we always appreciate a pull request to the repository, or a report on the issue tracker. Thanks!`

### Dropout function

Helps prevent overfitting by adding a random distortion to the input data during
training. Specifically, cells of the input are zeroed with probability
determined by the `dropout_rate`

argument. Cells which are not zeroed are
rescaled by `1-rate`

. When not in training mode, the distortion is disabled (see
Hinton et al., 2012).

`Example````
from thinc.api import chain, Linear, Dropout
model = chain(Linear(10, 2), Dropout(0.2))
Y, backprop = model(X, is_train=True)
# Configure dropout rate via the dropout_rate attribute.
for node in model.walk():
if node.name == "dropout":
node.attrs["dropout_rate"] = 0.5
```

Argument | Type | Description |
---|---|---|

`dropout_rate` | `float` | The probability of zeroing the activations (default: 0). Higher dropout rates mean more distortion. Values around `0.2` are often good. |

RETURNS | `Model[T, T]` | The created dropout layer. |

View on GitHub## thinc/layers/dropout.py

`Can't fetch code example from GitHub :( Please use the link above to view the example. If you've come across a broken link, we always appreciate a pull request to the repository, or a report on the issue tracker. Thanks!`

### Embed function

Map integers to vectors, using a fixed-size lookup table. The input to the layer should be a two-dimensional array of integers, one column of which the embeddings table will slice as the indices.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nV` | `int` | Number of input vectors. Defaults to `1` . |

keyword-only | ||

`column` | `int` | The column to slice from the input, to get the indices. |

`initializer` | `Optional[Callable]` | A function to initialize the internal parameters. Defaults to `uniform_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting (default `None` ). |

RETURNS | `Model[Union[Ints1d, Ints2d], Floats2d]` | The created embedding layer. |

View on GitHub## thinc/layers/embed.py

### HashEmbed function

An embedding layer that uses the “hashing trick” to map keys to distinct values. The hashing trick involves hashing each key four times with distinct seeds, to produce four likely differing values. Those values are modded into the table, and the resulting vectors summed to produce a single result. Because it’s unlikely that two different keys will collide on all four “buckets”, most distinct keys will receive a distinct vector under this scheme, even when the number of vectors in the table is very low.

Argument | Type | Description |
---|---|---|

`nO` | `int` | The size of the output vectors. |

`nV` | `int` | Number of input vectors. |

keyword-only | ||

`seed` | `Optional[int]` | A seed to use for the hashing. |

`column` | `int` | The column to select features from. |

`initializer` | `Optional[Callable]` | A function to initialize the internal parameters. Defaults to `uniform_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting (default `None` ). |

RETURNS | `Model[Union[Ints1d, Ints2d], Floats2d]` | The created embedding layer. |

View on GitHub## thinc/layers/hashembed.py

### LayerNorm function

Perform layer normalization on the inputs (Ba et al., 2016). This layer does not change the dimensionality of the vectors.

Argument | Type | Description |
---|---|---|

`nI` | `Optional[int]` | The size of the input vectors. |

RETURNS | `Model[Floats2d, Floats2d]` | The created normalization layer. |

View on GitHub## thinc/layers/layernorm.py

### Linear function

The `Linear`

layer multiplies inputs by a weights matrix `W`

and adds a bias
vector `b`

. In PyTorch this is called a `Linear`

layer, while Keras calls it a
`Dense`

layer.

`Example````
from thinc.api import Linear
model = Linear(10, 5)
model.initialize()
Y = model.predict(model.ops.alloc2f(2, 5))
assert Y.shape == (2, 10)
```

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Callable` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` . |

`init_b` | `Callable` | A function to initialize the bias vector. Defaults to `zero_init` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created `Linear` layer. |

View on GitHub## thinc/layers/linear.py

### Sigmoid function

A linear (aka dense) layer, followed by a sigmoid activation. This is usually
used as an output layer for multi-label classification (in contrast to the
`Softmax`

layer, which is used for problems where exactly one class is correct
per example.

Argument | Type | Description |
---|---|---|

`nOs` | `Tuple[int, …]` | The sizes of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

RETURNS | `Model[Floats2d, Floats2d]` | The created sigmoid layer. |

View on GitHub## thinc/layers/sigmoid.py

### sigmoid_activation function

Apply the sigmoid logistic function as an activation to the inputs. This is
often used as an output activation for multi-label classification, because each
element of the output vectors will be between `0`

and `1`

.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Floats2d, Floats2d]` | The created `sigmoid_activation` layer. |

View on GitHub## thinc/layers/sigmoid_activation.py

### LSTM and BiLSTM function

An LSTM recurrent neural network. The BiLSTM is bidirectional: that is, each
layer concatenated a forward LSTM with an LSTM running in the reverse direction.
If you are able to install PyTorch, you should usually prefer to use the
`PyTorchLSTM`

layer instead of Thinc’s implementations, as PyTorch’s LSTM
implementation is significantly faster.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`bi` | `bool` | Use BiLSTM. |

`depth` | `int` | Number of layers (default `1` ). |

`dropout` | `float` | Dropout rate to avoid overfitting (default `0` ). |

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

RETURNS | `Model[Padded, Padded]` | The created LSTM layer(s). |

View on GitHub## thinc/layers/lstm.py

### Maxout function

A dense layer with a “maxout” activation
(Goodfellow et al, 2013). Maxout layers
require a weights array of shape `(nO, nP, nI)`

in order to compute outputs of
width `nO`

given inputs of width `nI`

. The extra multiple, `nP`

, determines the
number of “pieces” that the piecewise-linear activation will consider.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

`nP` | `int` | Number of maxout pieces (default: 3). |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization, (default: False). |

RETURNS | `Model[Floats2d, Floats2d]` | The created maxout layer. |

View on GitHub## thinc/layers/maxout.py

### Mish function

A dense layer with Mish activation (Misra, 2019).

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization, (default: False). |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/mish.py

### Swish function

A dense layer with the Swish activation function
(Ramachandran et al., 2017). Swish is a
self-gating non-monotonic activation function similar to `GELU`

:
whereas GELU uses the CDF of the Gaussian distribution Φ for self-gating
`x * Φ(x)`

Swish uses the logistic CDF `x * σ(x)`

. Sometimes referred to as
“SiLU” for “Sigmoid Linear Unit”.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `he_normal_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/swish.py

### Gelu function

A dense layer with the GELU activation function
(Hendrycks and Gimpel, 2016). The GELU or
“Gaussian Error Linear Unit” is a self-gating non-monotonic activation function
similar to Swish: whereas GELU uses the CDF of the Gaussian
distribution Φ for self-gating `x * Φ(x)`

the Swish activation uses the logistic
CDF σ and computes `x * σ(x)`

. Various approximations exist, but `thinc`

implements the exact GELU. The use of GELU is popular within transformer
feed-forward blocks.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `he_normal_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/gelu.py

### ReluK function

A dense layer with the ReLU activation function where the maximum value is
clipped at `k`

. A common choice is `k=6`

introduced for convolutional deep
belief networks
(Krizhevsky, 2010).
The resulting function `relu6`

is commonly used in low-precision scenarios.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

`k` | `float` | Maximum value. Defaults to `6.0` .. |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/clipped_linear.py#L132

### HardSigmoid function

A dense layer with hard sigmoid activation function, which is a fast linear
approximation of sigmoid, defined as `max(0, min(1, x * 0.2 + 0.5))`

.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/clipped_linear.py#L90

### HardTanh function

A dense layer with hard tanh activation function, which is a fast linear
approximation of tanh, defined as `max(-1, min(1, x))`

.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/clipped_linear.py#L111

### ClippedLinear function

A dense layer implementing a flexible clipped linear activation function of the
form `max(min_value, min(max_value, x * slope + offset))`

. It is used to
implement the `ReluK`

, `HardSigmoid`

, and
`HardTanh`

layers.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

`slope` | `float` | The slope of the linear function: `input * slope` . |

`offset` | `float` | The offset or intercept of the linear function: `input * slope + offset` . |

`min_val` | `float` | Minimum value to clip to. |

`max_val` | `float` | Maximum value to clip to. |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/clipped_linear.py

### HardSwish function

A dense layer implementing the hard Swish activation function, which is a fast
linear approximation of Swish: `x * hard_sigmoid(x)`

.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `he_normal_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/hard_swish.py

### HardSwishMobileNet function

A dense layer implementing the a variant of the fast linear hard Swish
activation function used in `MobileNetV3`

(Howard et al., 2019), defined as
`x * (relu6(x + 3) / 6)`

.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `he_normal_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization. Defaults to `False` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created dense layer. |

View on GitHub## thinc/layers/hard_swish_mobilenet.py

### MultiSoftmax function

Neural network layer that predicts several multi-class attributes at once. For instance, we might predict one class with six variables, and another with five. We predict the 11 neurons required for this, and then softmax them such that columns 0-6 make a probability distribution and columns 6-11 make another.

Argument | Type | Description |
---|---|---|

`nOs` | `Tuple[int, …]` | The sizes of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

RETURNS | `Model[Floats2d, Floats2d]` | The created multi softmax layer. |

View on GitHub## thinc/layers/multisoftmax.py

### ParametricAttention function

A layer that uses the parametric attention scheme described by Yang et al. (2016). The layer learns a parameter vector that is used as the keys in a single-headed attention mechanism.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

RETURNS | `Model[Ragged, Ragged]` | The created attention layer. |

View on GitHub## thinc/layers/parametricattention.py

### Relu function

A dense layer with Relu activation.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `glorot_uniform_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`dropout` | `Optional[float]` | Dropout rate to avoid overfitting. |

`normalize` | `bool` | Whether or not to apply layer normalization, (default: False). |

RETURNS | `Model[Floats2d, Floats2d]` | The created Relu layer. |

View on GitHub## thinc/layers/relu.py

### Softmax function

A dense layer with a softmax activation. This is usually used as a prediction layer. Vectors produced by the softmax function sum to 1, and have values between 0 and 1, so each vector can be interpreted as a probability distribution.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `zero_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created softmax layer. |

View on GitHub## thinc/layers/softmax.py

### Softmax_v2 function

A dense layer with a softmax activation. This is usually used as a prediction layer. Vectors produced by the softmax function sum to 1, and have values between 0 and 1, so each vector can be interpreted as a probability distribution.

`Softmax_v2`

supports outputting unnormalized probabilities during inference by
using `normalize_outputs=False`

as an argument. This is useful when we are only
interested in finding the top-k classes, but not their probabilities. Computing
unnormalized probabilities is faster, because it skips the expensive
normalization step.

The `temperature`

argument of `Softmax_v2`

provides control of the softmax
distribution. Values larger than 1 increase entropy and values between 0 and 1
(exclusive) decrease entropy of the distribution. The default temperature of 1
will calculate the unmodified softmax distribution. `temperature`

is not used
during inference when `normalize_outputs=False`

.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`nI` | `Optional[int]` | The size of the input vectors. |

keyword-only | ||

`init_W` | `Optional[Callable]` | A function to initialize the weights matrix. Defaults to `zero_init` when set to `None` . |

`init_b` | `Optional[Callable]` | A function to initialize the bias vector. Defaults to `zero_init` when set to `None` . |

`normalize_outputs` | `bool` | Return normalized probabilities during inference. Defaults to `True` . |

`temperature` | `float` | Temperature to divide logits by. Defaults to `1.0` . |

RETURNS | `Model[Floats2d, Floats2d]` | The created softmax layer. |

View on GitHub## thinc/layers/softmax.py

### SparseLinear function

A sparse linear layer using the “hashing trick”. Useful for tasks such as text
classification. Inputs to the layer should be a tuple of arrays
`(keys, values, lengths)`

, where the `keys`

and `values`

are arrays of the same
length, describing the concatenated batch of input features and their values.
The `lengths`

array should have one entry per sequence in the batch, and the sum
of the lengths should equal the length of the keys and values array.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`length` | `int` | The size of the weights vector, to be tuned empirically. |

RETURNS | `Model[Tuple[ArrayXd, ArrayXd, ArrayXd], ArrayXd]` | The created layer. |

View on GitHub## thinc/layers/sparselinear.pyx

### SparseLinear_v2 functionNew: v8.1.6

A sparse linear layer using the “hashing trick”. Useful for tasks such as text
classification. Inputs to the layer should be a tuple of arrays
`(keys, values, lengths)`

, where the `keys`

and `values`

are arrays of the same
length, describing the concatenated batch of input features and their values.
The `lengths`

array should have one entry per sequence in the batch, and the sum
of the lengths should equal the length of the keys and values array.

Argument | Type | Description |
---|---|---|

`nO` | `Optional[int]` | The size of the output vectors. |

`length` | `int` | The size of the weights vector, to be tuned empirically. |

RETURNS | `Model[Tuple[ArrayXd, ArrayXd, ArrayXd], ArrayXd]` | The created layer. |

View on GitHub## thinc/layers/sparselinear.pyx

## Reduction operations

### reduce_first function

Pooling layer that reduces the dimensions of the data by selecting the first item of each sequence. This is most useful after multi-head attention layers, which can learn to assign a good feature representation for the sequence to one of its elements.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Ragged, ArrayXd]` | The created pooling layer. |

View on GitHub## thinc/layers/reduce_first.py

### reduce_last function

Pooling layer that reduces the dimensions of the data by selecting the last item of each sequence. This is typically used after multi-head attention or recurrent neural network layers such as LSTMs, which can learn to assign a good feature representation for the sequence to its final element.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Ragged, ArrayXd]` | The created pooling layer. |

View on GitHub## thinc/layers/reduce_last.py

### reduce_max function

Pooling layer that reduces the dimensions of the data by selecting the maximum
value for each feature. A `ValueError`

is raised if any element in `lengths`

is
zero.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Ragged, Floats2d]` | The created pooling layer. |

View on GitHub## thinc/layers/reduce_max.py

### reduce_mean function

Pooling layer that reduces the dimensions of the data by computing the average value of each feature. Zero-length sequences are reduced to the zero vector.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Ragged, Floats2d]` | The created pooling layer. |

View on GitHub## thinc/layers/reduce_mean.py

### reduce_sum function

Pooling layer that reduces the dimensions of the data by computing the sum for each feature. Zero-length sequences are reduced to the zero vector.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Ragged, Floats2d]` | The created pooling layer. |

View on GitHub## thinc/layers/reduce_sum.py

## Combinators

Combinators are layers that express **higher-order functions**: they take one or
more layers as arguments and express some relationship or perform some
additional logic around the child layers. Combinators can also be used to
overload operators. For example, binding `chain`

to `>>`

allows you to write `Relu(512) >> Softmax()`

instead of
`chain(Relu(512), Softmax())`

.

### add function

Compose two or more models `f`

, `g`

, etc, such that their outputs are added,
i.e. `add(f, g)(x)`

computes `f(x) + g(x)`

.

Argument | Type | Description |
---|---|---|

`*layers` | `Model[Any, ArrayXd]` | The models to compose. |

RETURNS | `Model[Any, ArrayXd]` | The composed model. |

View on GitHub## thinc/layers/add.py

### bidirectional function

Stitch two RNN models into a bidirectional layer. Expects squared sequences.

Argument | Type | Description |
---|---|---|

`l2r` | `Model[Padded, Padded]` | The first model. |

`r2l` | `Optional[Model[Padded, Padded]]` | The second model. |

RETURNS | `Model[Padded, Padded]` | The composed bidirectional layer. |

View on GitHub## thinc/layers/bidirectional.py

### chain function

Compose two or more models such that they become layers of a single feed-forward
model, e.g. `chain(f, g)`

computes `g(f(x))`

.

Argument | Type | Description |
---|---|---|

`layer1 ` | `Model` | The first model to compose. |

`layer2` | `Model` | The second model to compose. |

`*layers` | `Model` | Any additional models to compose. |

RETURNS | `Model` | The composed feed-forward model. |

View on GitHub## thinc/layers/chain.py

### clone function

Construct `n`

copies of a layer, with distinct weights. For example,
`clone(f, 3)(x)`

computes `f(f'(f''(x)))`

.

Argument | Type | Description |
---|---|---|

`orig` | `Model` | The layer to copy. |

`n` | `int` | The number of copies to construct. |

RETURNS | `Model` | A composite model containing two or more copies. |

View on GitHub## thinc/layers/clone.py

### concatenate function

Compose two or more models `f`

, `g`

, etc, such that their outputs are
concatenated, i.e. `concatenate(f, g)(x)`

computes `hstack(f(x), g(x))`

.

Argument | Type | Description |
---|---|---|

`*layers` | `Model` , … | The models to compose. |

RETURNS | `Model` | The composed model. |

View on GitHub## thinc/layers/concatenate.py

### map_list function

Map a child layer across list inputs.

Argument | Type | Description |
---|---|---|

`layer` | `Model[InT, OutT]` | The child layer to map. |

RETURNS | `Model[List[InT], List[OutT]]` | The composed model. |

View on GitHub## thinc/layers/map_list.py

### expand_window function

For each vector in an input, construct an output vector that contains the input
and a window of surrounding vectors. This is one step in a convolution. If the
`window_size`

is three, the output size `nO`

will be `nI * 7`

after
concatenating three contextual vectors from the left, and three from the right,
to each input vector. In general, `nO`

equals `nI * (2 * window_size + 1)`

.

Argument | Type | Description |
---|---|---|

`window_size` | `int` | The window size (default 1) that determines the number of surrounding vectors. |

RETURNS | `Model[T, T]` | The created layer for adding context to vectors. |

View on GitHub## thinc/layers/expand_window.py

### noop function

Transform a sequences of layers into a null operation.

Argument | Type | Description |
---|---|---|

`*layers` | `Model` | The models to compose. |

RETURNS | `Model` | The composed model. |

View on GitHub## thinc/layers/noop.py

### residual function

A unary combinator creating a residual connection. This converts a layer
computing `f(x)`

into one that computes `f(x)+x`

. Gradients flow through
residual connections directly, helping the network to learn more smoothly.

Argument | Type | Description |
---|---|---|

`layer` | `Model[T, T]` | A model with the same input and output types. |

RETURNS | `Model[T, T]` | A model with the unchanged input and output types. |

View on GitHub## thinc/layers/residual.py

### tuplify function

Give each child layer a separate copy of the input, and the combine the output of the child layers into a tuple. Useful for providing original and modified input to a downstream layer.

On the backward pass the loss from each child is added together, so when using custom datatypes they should define an addition operator.

Argument | Type | Description |
---|---|---|

`*layers` | `Model[Any, T] …` | The models to compose. |

RETURNS | `Model[Any, Tuple[T]]` | The composed feed-forward model. |

View on GitHub## thinc/layers/tuplify.py

### siamese function

Combine and encode a layer and a similarity function to form a siamese architecture. Typically used to learn symmetric relationships, such as redundancy detection.

Argument | Type | Description |
---|---|---|

`layer` | `Model` | The layer to run over the pair of inputs. |

`similarity` | `Model` | The similarity layer. |

RETURNS | `Model[Tuple, ArrayXd]` | The created siamese layer. |

View on GitHub## thinc/layers/siamese.py

### uniqued function

Group inputs to a layer, so that the layer only has to compute for the unique
values. The data is transformed back before output, and the same transformation
is applied for the gradient. Effectively, this is a cache local to each
minibatch. The `uniqued`

wrapper is useful for word inputs, because common words
are seen often, but we may want to compute complicated features for the words,
using e.g. character LSTM.

Argument | Type | Description |
---|---|---|

`layer` | `Model` | The layer. |

keyword-only | ||

`column` | `int` | The column. Defaults to `0` . |

RETURNS | `Model[Ints2d, Floats2d]` | The composed model. |

View on GitHub## thinc/layers/uniqued.py

## Data type transfers

### array_getitem, ints_getitem, floats_getitem function

Index into input arrays, and return the subarrays. Multi-dimensional indexing
can be performed by passing in a tuple, and slicing can be performed using the
slice object. For instance, `X[:, :-1]`

would be
`(slice(None, None), slice(None, -1))`

.

Argument | Type | Description |
---|---|---|

`index` | `Union[Union[int, slice, Sequence[int]], Tuple[Union[int, slice, Sequence[int]], …]` | A valid numpy-style index. |

View on GitHub## thinc/layers/array_getitem.py

### list2array function

Transform sequences to ragged arrays if necessary. If sequences are already
ragged, do nothing. A ragged array is a tuple `(data, lengths)`

, where `data`

is
the concatenated data.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[List2d, Array2d]` | The layer to compute the transformation. |

View on GitHub## thinc/layers/list2array.py

### list2ragged function

Transform sequences to ragged arrays if necessary and return the ragged array.
If sequences are already ragged, do nothing. A ragged array is a tuple
`(data, lengths)`

, where `data`

is the concatenated data.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[ListXd, Ragged]` | The layer to compute the transformation. |

View on GitHub## thinc/layers/list2ragged.py

### list2padded function

Create a layer to convert a list of array inputs into
`Padded`

.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[List2d, Padded]` | The layer to compute the transformation. |

View on GitHub## thinc/layers/list2padded.py

### ragged2list function

Transform sequences from a ragged format into lists.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Ragged, ListXd]` | The layer to compute the transformation. |

View on GitHub## thinc/layers/ragged2list.py

### padded2list function

Create a layer to convert a `Padded`

input into a list
of arrays.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Padded, List2d]` | The layer to compute the transformation. |

View on GitHub## thinc/layers/padded2list.py

### remap_ids function

Remap a sequence of strings, integers or other hashable inputs using a mapping
table, usually as a preprocessing step before embeddings. The input can also be
a two dimensional integer array in which case the `column`

attribute tells the
`remap_ids`

layer which column of the array to map with the `mapping_table`

.
Both attributes can be passed on initialization, but since the layer is designed
to retrieve them from `model.attrs`

during `forward`

, they can be set any time
before calling `forward`

. This means that they can also be changed between
calls. Before calling `forward`

the `mapping_table`

has to be set and for 2D
inputs the `column`

is also required.

Argument | Type | Description |
---|---|---|

`mapping_table` | `Dict[Any, int]` | The mapping table to use. Can also be set after initialization by writing to `model.attrs["mapping_table"]` . |

`default` | `int` | The default value if the input does not have an entry in the mapping table. |

`column` | `int` | The column to apply the mapper to in case of 2D input. |

RETURNS | `Model[Union[Sequence[Hashable], Ints1d, Ints2d], Ints2d]` | The layer to compute the transformation. |

View on GitHub## thinc/layers/remap_ids.py

### strings2arrays function

**Input:**`Sequence[Sequence[str]]`

**Output:**`List[Ints2d]`

Transform a sequence of string sequences to a list of arrays.

Argument | Type | Description |
---|---|---|

RETURNS | `Model[Sequence[Sequence[str]], List[Ints2d]]` | The layer to compute the transformation. |

View on GitHub## thinc/layers/strings2arrays.py

### with_array function

Transform sequence data into a contiguous array on the way into and out of a model. Handles a variety of sequence types: lists, padded and ragged. If the input is an array, it is passed through unchanged.

Argument | Type | Description |
---|---|---|

`layer` | `Model[ArrayXd, ArrayXd]` | The layer to wrap. |

keyword-only | ||

`pad` | `int` | The padding. Defaults to `0` . |

RETURNS | `Model` | The wrapped layer. |

View on GitHub## thinc/layers/with_array2d.py

### with_array2d function

Transform sequence data into a contiguous two-dimensional array on the way into
and out of a model. In comparison to the `with_array`

layer, the behavior of
this layer mostly differs on `Padded`

inputs, as this layer merges the batch and
length axes to form a two-dimensional array. Handles a variety of sequence
types: lists, padded and ragged. If the input is a two-dimensional array, it is
passed through unchanged.

Argument | Type | Description |
---|---|---|

`layer` | `Model[Array2d, Array2d]` | The layer to wrap. |

keyword-only | ||

`pad` | `int` | The padding. Defaults to `0` . |

RETURNS | `Model` | The wrapped layer. |

View on GitHub## thinc/layers/with_array.py

### with_flatten function

**Input:**`Sequence[Sequence[Any]]`

**Output:**`ListXd`

Flatten nested inputs on the way into a layer and reverse the transformation over the outputs.

Argument | Type | Description |
---|---|---|

`layer` | `Model[Sequence[Any], ArrayXd]` | The layer to wrap. |

RETURNS | `Model[Sequence[Sequence[Any]], ListXd]` | The wrapped layer. |

View on GitHub## thinc/layers/with_flatten.py

### with_flatten_v2 functionNew: v8.1.6

**Input:**`List[List[InItemT]]`

**Output:**`List[List[OutItemT]]`

Flatten nested inputs on the way into a layer and reverse the transformation over the outputs.

Argument | Type | Description |
---|---|---|

`layer` | `Model[List[InItemT], List[OutItemT]]` | The layer to wrap. |

RETURNS | `Model[List[List[InItemT]], List[List[OutItemT]]]` | The wrapped layer. |

View on GitHub## thinc/layers/with_flatten_v2.py

### with_padded function

Convert sequence input into the `Padded`

data type on
the way into a layer and reverse the transformation on the output.

Argument | Type | Description |
---|---|---|

`layer` | `Model[Padded, Padded]` | The layer to wrap. |

RETURNS | `Model` | The wrapped layer. |

View on GitHub## thinc/layers/with_padded.py

### with_ragged function

Convert sequence input into the `Ragged`

data type on
the way into a layer and reverse the transformation on the output.

Argument | Type | Description |
---|---|---|

`layer` | `Model[Ragged, Ragged]` | The layer to wrap. |

RETURNS | `Model` | The wrapped layer. |

View on GitHub## thinc/layers/with_ragged.py

### with_list function

Convert sequence input into lists on the way into a layer and reverse the transformation on the outputs.

Argument | Type | Description |
---|---|---|

`layer` | `Model[List2d, List2d]` | The layer to wrap. |

RETURNS | `Model` | The wrapped layer. |

View on GitHub## thinc/layers/with_list.py

### with_getitem function

**Input:**`Tuple`

**Output:**`Tuple`

Transform data on the way into and out of a layer by plucking an item from a tuple.

Argument | Type | Description |
---|---|---|

`idx` | `int` | The index to pluck from the tuple. |

`layer` | `Model[ArrayXd, ArrayXd]` | The layer to wrap. |

RETURNS | `Model[Tuple, Tuple]` | The wrapped layer. |

View on GitHub## thinc/layers/with_getitem.py

### with_reshape function

Reshape data on the way into and out from a layer.

Argument | Type | Description |
---|---|---|

`layer` | `Model[Array2d, Array2d]` | The layer to wrap. |

RETURNS | `Model[Array3d, Array3d]` | The wrapped layer. |

View on GitHub## thinc/layers/with_reshape.py

### with_debug function

**Input:**`Any`

**Output:**`Any`

Debugging layer that wraps any layer and allows executing callbacks during the forward pass, backward pass and initialization. The callbacks will receive the same arguments as the functions they’re called in and are executed before the function runs.

`Example````
from thinc.api import Linear, with_debug
def on_init(model, X, Y):
print(f"X: {type(Y)}, Y ({type(Y)})")
model = with_debug(Linear(2, 5), on_init=on_init)
model.initialize()
```

Argument | Type | Description |
---|---|---|

`layer` | `Model` | The layer to wrap. |

`name` | `Optional[str]` | Optional name for the wrapped layer, will be prefixed by `debug:` . Defaults to name of the wrapped layer. |

keyword-only | ||

`on_init` | `Callable[[Model, Any, Any], None]` | Function called on initialization. Receives the model and the `X` and `Y` passed to `Model.initialize` , if available. |

`on_forward` | `Callable[[Model, Any, bool], None]` | Function called at the start of the forward pass. Receives the model, the inputs and the value of `is_train` . |

`on_backprop` | `Callable[[Any], None] = do_nothing` | Function called at the start of the backward pass. Receives the gradient. |

RETURNS | `Model` | The wrapped layer. |

View on GitHub## thinc/layers/with_debug.py

### with_nvtx_range function

**Input:**`Any`

**Output:**`Any`

Layer that wraps any layer and marks the forward and backprop passes as an NVTX range. This can be helpful when profiling GPU performance of a layer.

`Example````
from thinc.api import Linear, with_nvtx_range
model = with_nvtx_range(Linear(2, 5))
model.initialize()
```

Argument | Type | Description |
---|---|---|

`layer` | `Model` | The layer to wrap. |

`name` | `Optional[str]` | Optional name for the wrapped layer. Defaults to the name of the wrapped layer. |

keyword-only | ||

`forward_color` | `int` | Identifier of the color to use for the forward pass |

`backprop_color` | `int` | Identifier of the color to use for the backward pass |

RETURNS | `Model` | The wrapped layer. |

View on GitHub## thinc/layers/with_nvtx_range.py

### with_signpost_interval functionNew: v8.1.1

**Input:**`Any`

**Output:**`Any`

Layer that wraps any layer and marks the init, forward and backprop passes as a
(macOS) signpost interval. This can be helpful when profiling the performance of
a layer using macOS
Instruments.app. Use of this
layer requires that the
`os-signpost`

package is installed.

`Example````
from os_signpost import Signposter
from thinc.api import Linear, with_signpost_interval
signposter = Signposter("com.example.my_subsystem",
Signposter.Category.DynamicTracing)
model = with_signpost_interval(Linear(2, 5), signposter)
model.initialize()
```

Argument | Type | Description |
---|---|---|

`layer` | `Model` | The layer to wrap. |

`signposter` | `os_signposter.Signposter` | `Signposter` object to log the interval with. |

`name` | `Optional[str]` | Optional name for the wrapped layer. Defaults to the name of the wrapped layer. |

RETURNS | `Model` | The wrapped layer. |

View on GitHub## thinc/layers/with_signpost_interval.py

## Wrappers

### PyTorchWrapper, PyTorchRNNWrapper function

**Input:**`Any`

**Output:**`Any`

Wrap a PyTorch model so that it has the same API as Thinc
models. To optimize the model, you’ll need to create a PyTorch optimizer and
call `optimizer.step`

after each batch. The `PyTorchRNNWrapper`

has the same
signature as the `PyTorchWrapper`

and lets you to pass in a custom sequence
model that has the same inputs and output behavior as a
`torch.nn.RNN`

object.

Your PyTorch model’s forward method can take arbitrary positional arguments and
keyword arguments, but must return either a **single tensor** as output or a
**tuple**. You may find
PyTorch’s `register_forward_hook`

helpful if you need to adapt the output. The convert functions are used to map
inputs and outputs to and from your PyTorch model. Each function should return
the converted output, and a callback to use during the backward pass:

```
Xtorch, get_dX = convert_inputs(X)
Ytorch, torch_backprop = model.shims[0](Xtorch, is_train)
Y, get_dYtorch = convert_outputs(Ytorch)
```

To allow maximum flexibility, the `PyTorchShim`

expects
`ArgsKwargs`

objects on the way into the forward
and backward passes. The `ArgsKwargs`

objects will be passed straight into the
model in the forward pass, and straight into `torch.autograd.backward`

during
the backward pass.

Argument | Type | Description |
---|---|---|

`pytorch_model` | `Any` | The PyTorch model. |

`convert_inputs` | `Callable` | Function to convert inputs to PyTorch tensors (same signature as `forward` function). |

`convert_outputs` | `Callable` | Function to convert outputs from PyTorch tensors (same signature as `forward` function). |

RETURNS | `Model[Any, Any]` | The Thinc model. |

View on GitHub## thinc/layers/pytorchwrapper.py

### TorchScriptWrapper_v1 functionNew: v8.1.6

**Input:**`Any`

**Output:**`Any`

Wrap a TorchScript model so that it
has the same API as Thinc models. To optimize the model, you’ll need to create a
PyTorch optimizer and call `optimizer.step`

after each batch.

Your TorchScript model’s forward method can take arbitrary positional arguments
and keyword arguments, but must return either a **single tensor** as output or a
**tuple**. The convert functions are used to map inputs and outputs to and from
your TorchScript model. Each function should return the converted output, and a
callback to use during the backward pass:

```
Xtorch, get_dX = convert_inputs(X)
Ytorch, torch_backprop = model.shims[0](Xtorch, is_train)
Y, get_dYtorch = convert_outputs(Ytorch)
```

To allow maximum flexibility, the `TorchScriptShim`

expects `ArgsKwargs`

objects on the way into the
forward and backward passes. The `ArgsKwargs`

objects will be passed straight
into the model in the forward pass, and straight into `torch.autograd.backward`

during the backward pass.

Note that the `torchscript_model`

argument can be `None`

. This is useful for
deserialization since serialized TorchScript contains both the model and its
weights.

A PyTorch wrapper can be converted to a TorchScript wrapper using the
`pytorch_to_torchscript_wrapper`

function:

```
from thinc.api import PyTorchWrapper_v2, pytorch_to_torchscript_wrapper
import torch
model = PyTorchWrapper_v2(torch.nn.Linear(nI, nO)).initialize()
script_model = pytorch_to_torchscript_wrapper(model)
```

Argument | Type | Description |
---|---|---|

`torchscript_model` | `Optional[torch.jit.ScriptModule]` | The TorchScript model. |

`convert_inputs` | `Callable` | Function to convert inputs to PyTorch tensors (same signature as `forward` function). |

`convert_outputs` | `Callable` | Function to convert outputs from PyTorch tensors (same signature as `forward` function). |

`mixed_precision` | `bool` | Enable mixed-precision training. |

`grad_scaler` | `Optional[PyTorchGradScaler]` | Gradient scaler to use during mixed-precision training. |

`device` | `Optional[torch.Device]` | The Torch device to execute the model on. |

RETURNS | `Model[Any, Any]` | The Thinc model. |

View on GitHub## thinc/layers/torchscriptwrapper.py

### TensorFlowWrapper function

**Input:**`Any`

**Output:**`Any`

Wrap a TensorFlow model, so that it has the same API
as Thinc models. To optimize the model, you’ll need to create a TensorFlow
optimizer and call `optimizer.apply_gradients`

after each batch. To allow
maximum flexibility, the `TensorFlowShim`

expects
`ArgsKwargs`

objects on the way into the forward
and backward passes.

Argument | Type | Description |
---|---|---|

`tensorflow_model` | `Any` | The TensorFlow model. |

RETURNS | `Model[Any, Any]` | The Thinc model. |

View on GitHub## thinc/layers/tensorflowwrapper.py

### MXNetWrapper function

**Input:**`Any`

**Output:**`Any`

Wrap a MXNet model, so that it has the same API as
Thinc models. To optimize the model, you’ll need to create a MXNet optimizer and
call `optimizer.step()`

after each batch. To allow maximum flexibility, the
`MXNetShim`

expects
`ArgsKwargs`

objects on the way into the forward
and backward passes.

Argument | Type | Description |
---|---|---|

`tensorflow_model` | `Any` | The TensorFlow model. |

RETURNS | `Model[Any, Any]` | The Thinc model. |

View on GitHub## thinc/layers/mxnetwrapper.py