Tensors are the core datastructure of TensorFlow.js They are a generalization of vectors and matrices to potentially higher dimensions.
We have utility functions for common cases like Scalar, 1D, 2D, 3D and 4D tensors, as well a number of functions to initialize tensors in ways useful for machine learning.
Creates a tf.Tensor with the provided values, shape and dtype.
// Pass an array of values to create a vector.
tf.tensor([1, 2, 3, 4]).print();
// Pass a nested array of values to make a matrix or a higher
// dimensional tensor.
tf.tensor([[1, 2], [3, 4]]).print();
// Pass a flat array and specify a shape yourself.
tf.tensor([1, 2, 3, 4], [2, 2]).print();
- values (TypedArray|Array) The values of the tensor. Can be nested array of numbers, or a flat array, or a TypedArray.
-
shape
(number[])
The shape of the tensor. Optional. If not provided,
it is inferred from
values
. Optional - dtype ('float32'|'int32'|'bool') The data type. Optional
Creates rank-0 tf.Tensor (scalar) with the provided value and dtype.
The same functionality can be achieved with tf.tensor(), but in general we recommend using tf.scalar() as it makes the code more readable.
tf.scalar(3.14).print();
- value (number|boolean) The value of the scalar.
- dtype ('float32'|'int32'|'bool') The data type. Optional
Creates rank-1 tf.Tensor with the provided values, shape and dtype.
The same functionality can be achieved with tf.tensor(), but in general we recommend using tf.tensor1d() as it makes the code more readable.
tf.tensor1d([1, 2, 3]).print();
- values (TypedArray|Array) The values of the tensor. Can be array of numbers, or a TypedArray.
- dtype ('float32'|'int32'|'bool') The data type. Optional
Creates rank-2 tf.Tensor with the provided values, shape and dtype.
The same functionality can be achieved with tf.tensor(), but in general we recommend using tf.tensor2d() as it makes the code more readable.
// Pass a nested array.
tf.tensor2d([[1, 2], [3, 4]]).print();
// Pass a flat array and specify a shape.
tf.tensor2d([1, 2, 3, 4], [2, 2]).print();
- values (TypedArray|Array) The values of the tensor. Can be nested array of numbers, or a flat array, or a TypedArray.
-
shape
([number, number])
The shape of the tensor. If not provided, it is inferred from
values
. Optional - dtype ('float32'|'int32'|'bool') The data type. Optional
Creates rank-3 tf.Tensor with the provided values, shape and dtype.
The same functionality can be achieved with tf.tensor(), but in general we recommend using tf.tensor3d() as it makes the code more readable.
// Pass a nested array.
tf.tensor3d([[[1], [2]], [[3], [4]]]).print();
// Pass a flat array and specify a shape.
tf.tensor3d([1, 2, 3, 4], [2, 2, 1]).print();
- values (TypedArray|Array) The values of the tensor. Can be nested array of numbers, or a flat array, or a TypedArray.
-
shape
([number, number, number])
The shape of the tensor. If not provided, it is inferred from
values
. Optional - dtype ('float32'|'int32'|'bool') The data type. Optional
Creates rank-4 tf.Tensor with the provided values, shape and dtype.
The same functionality can be achieved with tf.tensor(), but in general we recommend using tf.tensor4d() as it makes the code more readable.
// Pass a nested array.
tf.tensor4d([[[[1], [2]], [[3], [4]]]]).print();
// Pass a flat array and specify a shape.
tf.tensor4d([1, 2, 3, 4], [1, 2, 2, 1]).print();
- values (TypedArray|Array) The values of the tensor. Can be nested array of numbers, or a flat array, or a TypedArray.
-
shape
([number, number, number, number])
The shape of the tensor. Optional. If not provided,
it is inferred from
values
. Optional - dtype ('float32'|'int32'|'bool') The data type. Optional
Creates an empty tf.TensorBuffer with the specified shape
and dtype
.
The values are stored in cpu as TypedArray. Fill the buffer using
buffer.set()
, or by modifying directly buffer.values
. When done,
call buffer.toTensor()
to get an immutable tf.Tensor with those values.
When done, call buffer.toTensor()
to get an immutable tf.Tensor with those
values.
// Create a buffer and set values at particular indices.
const buffer = tf.buffer([2, 2]);
buffer.set(3, 0, 0);
buffer.set(5, 1, 0);
// Convert the buffer back to a tensor.
buffer.toTensor().print();
- shape (number[]) An array of integers defining the output tensor shape.
- dtype ('float32'|'int32'|'bool') The dtype of the buffer. Defaults to 'float32'. Optional
- values (TypedArray) The values of the buffer as TypedArray. Defaults to zeros. Optional
Creates a new tensor with the same values and shape as the specified tensor.
const x = tf.tensor([1, 2]);
x.clone().print();
- x (tf.Tensor) The tensor to clone.
Creates a tf.Tensor filled with a scalar value.
tf.fill([2, 2], 4).print();
- shape (number[]) An array of integers defining the output tensor shape.
- value (number) The scalar value to fill the tensor with.
- dtype ('float32'|'int32'|'bool') The type of an element in the resulting tensor. Defaults to 'float'. Optional
Creates a tf.Tensor from an image.
const image = new ImageData(1, 1);
image.data[0] = 100;
image.data[1] = 150;
image.data[2] = 200;
image.data[3] = 255;
tf.fromPixels(image).print();
- pixels (ImageData|HTMLImageElement|HTMLCanvasElement|HTMLVideoElement) The input image to construct the tensor from.
- numChannels (number) The number of channels of the output tensor. The supported image types are all 4-channel by default, a numChannels value less than 4 allows you to ignore channels. Optional
Return an evenly spaced sequence of numbers over the given interval.
tf.linspace(0, 9, 10).print();
- start (number) The start value of the sequence.
- stop (number) The end value of the sequence.
- num (number) The number of values to generate.
Creates a one-hot tf.Tensor. The locations represented by indices
take
value onValue
(defaults to 1), while all other locations take value
offValue
(defaults to 0).
tf.oneHot(tf.tensor1d([0, 1]), 3).print();
- indices (tf.Tensor1D) 1D Array of indices.
- depth (number) The depth of the one hot dimension.
- onValue (number) A number used to fill in output when the index matches the location. Optional
- offValue (number) A number used to fill in the output when the index does not match the location. Optional
Creates a tf.Tensor with all elements set to 1.
tf.ones([2, 2]).print();
- shape (number[]) An array of integers defining the output tensor shape.
- dtype ('float32'|'int32'|'bool') The type of an element in the resulting tensor. Defaults to 'float'. Optional
Creates a tf.Tensor with all elements set to 1 with the same shape as the given tensor.
const x = tf.tensor([1, 2]);
tf.onesLike(x).print();
- x (tf.Tensor) A tensor.
Prints information about the tf.Tensor including its data.
const verbose = true;
tf.tensor2d([1, 2, 3, 4], [2, 2]).print(verbose);
- x (tf.Tensor) The tensor to be printed.
-
verbose
(boolean)
Whether to print verbose information about the
Tensor
, including dtype and size. Optional
Creates a tf.Tensor with values sampled from a normal distribution.
tf.randomNormal([2, 2]).print();
- shape (number[]) An array of integers defining the output tensor shape.
- mean (number) The mean of the normal distribution. Optional
- stdDev (number) The standard deviation of the normal distribution. Optional
- dtype ('float32'|'int32') The data type of the output. Optional
- seed (number) The seed for the random number generator. Optional
Creates a tf.Tensor with values sampled from a uniform distribution.
The generated values follow a uniform distribution in the range [minval, maxval). The lower bound minval is included in the range, while the upper bound maxval is excluded.
tf.randomUniform([2, 2]).print();
- shape (number[]) An array of integers defining the output tensor shape.
- minval (number) The lower bound on the range of random values to generate. Defaults to 0. Optional
- maxval (number) The upper bound on the range of random values to generate. Defaults to 1. Optional
- dtype ('float32'|'int32'|'bool') The data type of the output tensor. Defaults to 'float32'. Optional
Creates a new tf.Tensor1D filled with the numbers in the range provided.
The tensor is a is half-open interval meaning it includes start, but excludes stop. Decrementing ranges and negative step values are also supported.
tf.range(0, 9, 2).print();
- start (number) An integer start value
- stop (number) An integer stop value
- step (number) An integer increment (will default to 1 or -1) Optional
- dtype ('float32'|'int32') The data type of the output tensor. Defaults to 'float32'. Optional
Creates a tf.Tensor with values sampled from a truncated normal distribution.
tf.truncatedNormal([2, 2]).print();
The generated values follow a normal distribution with specified mean and standard deviation, except that values whose magnitude is more than 2 standard deviations from the mean are dropped and re-picked.
- shape (number[]) An array of integers defining the output tensor shape.
- mean (number) The mean of the normal distribution. Optional
- stdDev (number) The standard deviation of the normal distribution. Optional
- dtype ('float32'|'int32') The data type of the output tensor. Optional
- seed (number) The seed for the random number generator. Optional
Creates a new variable with the provided initial value.
const x = tf.variable(tf.tensor([1, 2, 3]));
x.assign(tf.tensor([4, 5, 6]));
x.print();
- initialValue (tf.Tensor) Initial value for the tensor.
- trainable (boolean) If true, optimizers are allowed to update it. Optional
- name (string) Name of the variable. Defaults to a unique id. Optional
- dtype ('float32'|'int32'|'bool') If set, initialValue will be converted to the given type. Optional
Creates a tf.Tensor with all elements set to 0.
tf.zeros([2, 2]).print();
- shape (number[]) An array of integers defining the output tensor shape.
- dtype ('float32'|'int32'|'bool') The type of an element in the resulting tensor. Can be 'float32', 'int32' or 'bool'. Defaults to 'float'. Optional
Creates a tf.Tensor with all elements set to 0 with the same shape as the given tensor.
const x = tf.tensor([1, 2]);
tf.zerosLike(x).print();
- x (tf.Tensor) The tensor of required shape.
This section shows the main Tensor related classes in TensorFlow.js and the methods we expose on them.
A tf.Tensor object represents an immutable, multidimensional array of numbers that has a shape and a data type.
See tf.tensor() for details on how to create a tf.Tensor.
Converts a tf.Tensor to a tf.Tensor2D.
- rows (number) Number of rows in tf.Tensor2D.
- columns (number) Number of columns in tf.Tensor2D.
Converts a tf.Tensor to a tf.Tensor3D.
- rows (number) Number of rows in tf.Tensor3D.
- columns (number) Number of columns in tf.Tensor3D.
- depth (number) Depth of tf.Tensor3D.
Converts a tf.Tensor to a tf.Tensor4D.
- rows (number) Number of rows in tf.Tensor4D.
- columns (number) Number of columns in tf.Tensor4D.
- depth (number) Depth of tf.Tensor4D.
- depth2 (number) 4th dimension of tf.Tensor4D.
Casts a tf.Tensor to a specified dtype.
- dtype ('float32'|'int32'|'bool') Data-type to cast the tensor to.
Returns a tf.TensorBuffer that holds the underlying data.
Asynchronously downloads the values from the tf.Tensor. Returns a promise of TypedArray that resolves when the computation has finished.
Synchronously downloads the values from the tf.Tensor. This blocks the UI thread until the values are ready, which can cause performance issues.
Prints the tf.Tensor. See tf.print() for details.
- verbose (boolean) Whether to print verbose information about the tensor, including dtype and size. Optional
Reshapes the tensor into the provided shape. See tf.reshape() for more details.
- newShape (number[]) An array of integers defining the output tensor shape.
Reshapes the tensor into the shape of the provided tensor.
- x (tf.Tensor) The tensor of required shape.
Returns a tf.Tensor that has expanded rank, by inserting a dimension into the tensor's shape. See tf.expandDims() for details.
- axis (number) The dimension index at which to insert shape of 1. Defaults to 0 (the first dimension). Optional
Returns a tf.Tensor with dimensions of size 1 removed from the shape. See tf.squeeze() for more details.
- axis (number[]) A list of numbers. If specified, only squeezes the dimensions listed. The dimension index starts at 0. It is an error to squeeze a dimension that is not 1. Optional
A mutable tf.Tensor, useful for persisting state, e.g. for training.
A mutable object, similar to tf.Tensor, that allows users to set values at locations before converting to an immutable tf.Tensor.
See tf.buffer() for creating a tensor buffer.
Sets a value in the buffer at a given location.
- value (number) The value to set.
- locs (number[]) The location indices.
Returns the value in the buffer at the provided location.
- locs (number[]) The location indices.
This section describes some common Tensor transformations for reshaping and type-casting.
Casts a tf.Tensor to a new dtype.
const x = tf.tensor1d([1.5, 2.5, 3]);
tf.cast(x, 'int32').print();
- x (tf.Tensor) The input tensor to be casted.
- dtype ('float32'|'int32'|'bool') The dtype to cast the input tensor to.
Returns a tf.Tensor that has expanded rank, by inserting a dimension into the tensor's shape.
const x = tf.tensor1d([1, 2, 3, 4]);
const axis = 1;
x.expandDims(axis).print();
- x (tf.Tensor) The input tensor whose dimensions to be expanded.
-
axis
(number)
The dimension index at which to insert shape of
1
. Defaults to 0 (the first dimension). Optional
Pads a tf.Tensor with a given value and paddings.
This operation currently only implements the CONSTANT
mode.
const x = tf.tensor1d([1, 2, 3, 4]);
x.pad([[1, 2]]).print();
- x (tf.Tensor) The tensor to pad.
-
paddings
(Array)
An array of length
R
(the rank of the tensor), where each element is a length-2 tuple of ints[padBefore, padAfter]
, specifying how much to pad along each dimension of the tensor. - constantValue (number) The pad value to use. Defaults to 0. Optional
Reshapes a tf.Tensor to a given shape.
Given a input tensor, returns a new tensor with the same values as the
input tensor with shape shape
.
If one component of shape is the special value -1, the size of that dimension is computed so that the total size remains constant. In particular, a shape of [-1] flattens into 1-D. At most one component of shape can be -1.
If shape is 1-D or higher, then the operation returns a tensor with shape shape filled with the values of tensor. In this case, the number of elements implied by shape must be the same as the number of elements in tensor.
const x = tf.tensor1d([1, 2, 3, 4]);
x.reshape([2, 2]).print();
- x (tf.Tensor) The input tensor to be reshaped.
- shape (number[]) An array of integers defining the output tensor shape.
Removes dimensions of size 1 from the shape of a tf.Tensor.
const x = tf.tensor([1, 2, 3, 4], [1, 1, 4]);
x.squeeze().print();
- x (tf.Tensor) The input tensor to be squeezed.
- axis (number[]) An optional list of numbers. If specified, only squeezes the dimensions listed. The dimension index starts at 0. It is an error to squeeze a dimension that is not 1. Optional
TensorFlow.js provides several operations to slice or extract parts of a tensor, or join multiple tensors together.
Concatenates a list of tf.Tensors along a given axis.
The tensors ranks and types must match, and their sizes must match in all
dimensions except axis
.
const a = tf.tensor1d([1, 2]);
const b = tf.tensor1d([3, 4]);
a.concat(b).print(); // or a.concat(b)
const a = tf.tensor1d([1, 2]);
const b = tf.tensor1d([3, 4]);
const c = tf.tensor1d([5, 6]);
tf.concat([a, b, c]).print();
const a = tf.tensor2d([[1, 2], [10, 20]]);
const b = tf.tensor2d([[3, 4], [30, 40]]);
const axis = 1;
tf.concat([a, b], axis).print();
- tensors (tf.Tensor[]) A list of tensors to concatenate.
- axis (number) The axis to concate along. Defaults to 0 (the first dim). Optional
Gather slices from tensor x
's axis axis
according to indices
.
const x = tf.tensor1d([1, 2, 3, 4]);
const indices = tf.tensor1d([1, 3, 3]);
x.gather(indices).print();
const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const indices = tf.tensor1d([1, 1, 0]);
x.gather(indices).print();
- x (tf.Tensor) The input tensor whose slices to be gathered.
- indices (tf.Tensor1D) The indices of the values to extract.
- axis (number) The axis over which to select values. Defaults to 0. Optional
Reverses a tf.Tensor along a specified axis.
const x = tf.tensor1d([1, 2, 3, 4]);
x.reverse().print();
const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const axis = 1;
x.reverse(axis).print();
- x (tf.Tensor) The input tensor to be reversed.
- axis (number|number[]) The set of dimensions to reverse. Must be in the range [-rank(x), rank(x)). Defaults to all axes. Optional
Extracts a slice from a tf.Tensor starting at coordinates begin
and is of size size
.
Also available are stricter rank-specific methods with the same signature
as this method that assert that x
is of the given rank:
tf.slice1d
tf.slice2d
tf.slice3d
tf.slice4d
const x = tf.tensor1d([1, 2, 3, 4]);
x.slice([1], [2]).print();
const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
x.slice([1, 0], [1, 2]).print();
Stacks a list of rank-R
tf.Tensors into one rank-(R+1)
tf.Tensor.
const a = tf.tensor1d([1, 2]);
const b = tf.tensor1d([3, 4]);
const c = tf.tensor1d([5, 6]);
tf.stack([a, b, c]).print();
- tensors (tf.Tensor[]) A list of tensor objects with the same shape and dtype.
- axis (number) The axis to stack along. Defaults to 0 (the first dim). Optional
Construct an tensor by repeating it the number of times given by reps.
This operation creates a new tensor by replicating tf.input() reps
times. The output tensor's i'th dimension has input.shape[i] * reps[i]
elements, and the values of tf.input() are replicated
reps[i]
times along the i'th dimension. For example, tiling
[a, b, c, d]
by [2]
produces [a, b, c, d, a, b, c, d]
.
const a = tf.tensor1d([1, 2]);
a.tile([2]).print(); // or a.tile([2])
const a = tf.tensor2d([1, 2, 3, 4], [2, 2]);
a.tile([1, 2]).print(); // or a.tile([1, 2])
- x (tf.Tensor) The tensor to transpose.
- reps (number[]) Determines the number of replications per dimension.
Models are one of the primary abstractions used in TensorFlow.js Layers. Models can be trained, evaluated, and used for prediction. A model's state (topology, and optionally, trained weights) can be restored from various formats.
Models are a collection of Layers, see Model Creation for for details about how Layers can be connected.
There are two primary ways of creating models.
- Sequential — Easiest, works if the models is a simple stack of each layer's input resting on the top of the previous layer's output.
- Model — Offers more control if the layers need to be wired together in graph-like ways — multiple 'towers', layers that skip a layer, etc.
Creates a tf.Sequential model. A sequential model is any model where the outputs of one layer are the inputs to the next layer, i.e. the model topology is a simple 'stack' of layers, with no branching or skipping.
This means that the first layer passed to a Sequential model should have a
defined input shape. What that means is that it should have received an
inputShape
or batchInputShape
argument, or for some type of layers
(recurrent, Dense...) an inputDim
argument.
The key difference between tf.model() and tf.sequential() is that tf.sequential() is less generic, supporting only a linear stack of layers. tf.model() is more generic and supports an arbitrary graph (without cycles) of layers.
Examples:
const model = tf.sequential();
// First layer must have a defined input shape
model.add(tf.layers.dense({units: 32, inputShape: [50]}));
// Afterwards, TF.js does automatic shape inference.
model.add(tf.layers.dense({units: 4}));
// Inspect the inferred shape of the model's output, which equals
// `[null, 4]`. The 1st dimension is the undetermined batch dimension; the
// 2nd is the output size of the model's last layer.
console.log(model.outputs[0].shape);
It is also possible to specify a batch size (with potentially undetermined
batch dimension, denoted by "null") for the first layer using the
batchInputShape
key. The following example is equivalent to the above:
const model = tf.sequential();
// First layer must have a defined input shape
model.add(tf.layers.dense({units: 32, batchInputShape: [null, 50]}));
// Afterwards, TF.js does automatic shape inference.
model.add(tf.layers.dense({units: 4}));
// Inspect the inferred shape of the model's output.
console.log(model.outputs[0].shape);
You can also use an Array
of already-constructed Layer
s to create
a tf.Sequential model:
const model = tf.sequential({
layers: [tf.layers.dense({units: 32, inputShape: [50]}),
tf.layers.dense({units: 4})]
});
console.log(model.outputs[0].shape);
- config (Object) Optional
- layers (tf.layers.Layer[]) Stack of layers for the model. Optional
- name (string) The name of this model. Optional
A model is a data structure that consists of Layers
and defines inputs
and outputs.
The key difference between tf.model() and tf.sequential() is that tf.model() is more generic, supporting an arbitrary graph (without cycles) of layers. tf.sequential() is less generic and supports only a linear stack of layers.
When creating a tf.Model, specify its input(s) and output(s). Layers are used to wire input(s) to output(s).
For example, the following code snippet defines a model consisting of
two dense
layers, with 10 and 4 units, respectively.
// Define input, which has a size of 5 (not including batch dimension).
const input = tf.input({shape: [5]});
// First dense layer uses relu activation.
const denseLayer1 = tf.layers.dense({units: 10, activation: 'relu'});
// Second dense layer uses softmax activation.
const denseLayer2 = tf.layers.dense({units: 2, activation: 'softmax'});
// Obtain the output symbolic tensor by applying the layers on the input.
const output = denseLayer2.apply(denseLayer1.apply(input));
// Create the model based on the inputs.
const model = tf.model({inputs: input, outputs: output});
// The model can be used for training, evaluation and prediction.
// For example, the following line runs prediction with the model on
// some fake data.
model.predict(tf.ones([2, 5])).print();
See also: tf.sequential(), tf.loadModel().
- config (Object)
- inputs (tf.SymbolicTensor|tf.SymbolicTensor[])
- outputs (tf.SymbolicTensor|tf.SymbolicTensor[])
- name (string) Optional
Used to instantiate an input to a model as a tf.SymbolicTensor.
Users should call the tf.input() factory function for consistency with other generator functions.
Example:
// Defines a simple logistic regression model with 32 dimensional input
// and 3 dimensional output.
const x = tf.input({shape: [32]});
const y = tf.layers.dense({units: 3, activation: 'softmax'}).apply(x);
const model = tf.model({inputs: x, outputs: y});
model.predict(tf.ones([2, 32])).print();
Note: tf.input() is only necessary when using tf.model(). When using
tf.sequential(), specify inputShape
for the first layer or use inputLayer
as the first layer.
- config (InputConfig)
-
shape
(number[])
A shape, not including the batch size. For instance,
shape=[32]
indicates that the expected input will be batches of 32-dimensional vectors. Optional -
batchShape
(number[])
A shape tuple (integer), including the batch size. For instance,
batchShape=[10, 32]
indicates that the expected input will be batches of 10 32-dimensional vectors.batchShape=[null, 32]
indicates batches of an arbitrary number of 32-dimensional vectors. Optional - name (string) An optional name string for the layer. Should be unique in a model (do not reuse the same name twice). It will be autogenerated if it isn't provided. Optional
- dtype ('float32'|'int32'|'bool') Optional
- sparse (boolean) A boolean specifying whether the placeholder to be created is sparse. Optional
Load a model, including its topology and optionally weights. See the Tutorial named "How to import a Keras Model" for usage examples.
-
modelConfigPath
(string)
A path to the
ModelAndWeightsConfig
JSON describing the model in the canonical TensorFlow.js format.The content of the JSON file is assumed to be a JSON object with the following fields and values:
- 'modelTopology': A JSON object that can be either of:
- a model architecture JSON consistent with the format of the return
value of
keras.Model.to_json()
- a full model JSON in the format of
keras.models.save_model()
.
- 'weightsManifest': A TensorFlow.js weights manifest.
See the Python converter function
save_model()
for more details.It is also assumed that model weights can be accessed from relative paths described by the
paths
fields in weights manifest.
A tf.Model is a directed, acyclic graph of Layer
s plus methods for
training, evaluation, prediction and saving.
tf.Model is the basic unit of training, inference and evaluation in TensorFlow.js. To create a tf.Model, use tf.model().
See also: tf.Sequential, tf.loadModel().
Configures and prepares the model for training and evaluation. Compiling
outfits the model with an optimizer, loss, and/or metrics. Calling fit
or evaluate
on an un-compiled model will throw an error.
-
config
(Object)
a
ModelCompileConfig
specifying the loss, optimizer, and metrics to be used for fitting and evaluating this model. -
optimizer
(string|tf.train.Optimizer)
An instance of
tf.train.Optimizer
or a string name for an Optimizer. - loss (string|string[]|{[outputName: string]: string}) String (name of objective function) or objective function. If the model has multiple outputs, you can use a different loss on each output by passing a dictionary or an Array of losses. The loss value that will be minimized by the model will then be the sum of all individual losses.
-
metrics
(string[]|{[outputName: string]: string})
List of metrics to be evaluated by the model during training and testing.
Typically you will use
metrics=['accuracy']
. To specify different metrics for different outputs of a multi-output model, you could also pass a dictionary. Optional
Returns the loss value & metrics values for the model in test mode.
Loss and metrics are specified during compile()
, which needs to happen
before calls to evaluate()
.
Computation is done in batches.
const model = tf.sequential({
layers: [tf.layers.dense({units: 1, inputShape: [10]})]
});
model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
const result = model.evaluate(
tf.ones([8, 10]), tf.ones([8, 1]), {batchSize: 4});
result.print();
-
x
(tf.Tensor|tf.Tensor[])
tf.Tensor of test data, or an
Array
of tf.Tensors if the model has multiple inputs. -
y
(tf.Tensor|tf.Tensor[])
tf.Tensor of target data, or an
Array
of tf.Tensors if the model has multiple outputs. -
config
(Object)
A
ModelEvaluateConfig
, containing optional fields. Optional - batchSize (number) Batch size (Integer). If unspecified, it will default to 32. Optional
- verbose (ModelLoggingVerbosity) Verbosity mode. Optional
- sampleWeight (tf.Tensor) Tensor of weights to weight the contribution of different samples to the loss and metrics. Optional
-
steps
(number)
integer: total number of steps (batches of samples)
before declaring the evaluation round finished. Ignored with the default
value of
undefined
. Optional
Generates output predictions for the input samples.
Computation is done in batches.
Note: the "step" mode of predict() is currently not supported. This is because the TensorFlow.js core backend is imperative only.
const model = tf.sequential({
layers: [tf.layers.dense({units: 1, inputShape: [10]})]
});
model.predict(tf.ones([8, 10]), {batchSize: 4}).print();
-
x
(tf.Tensor|tf.Tensor[])
The input data, as an Tensor, or an
Array
of tf.Tensors if the model has multiple inputs. -
config
(Object)
A
ModelPredictConfig
object containing optional fields. Optional - batchSize (number) Batch size (Integer). If unspecified, it will default to 32. Optional
- verbose (boolean) Verbosity mode. Defaults to false. Optional
Returns predictions for a single batch of samples.
const model = tf.sequential({
layers: [tf.layers.dense({units: 1, inputShape: [10]})]
});
model.predictOnBatch(tf.ones([8, 10])).print();
- x (tf.Tensor) : Input samples, as an Tensor
Trains the model for a fixed number of epochs (iterations on a dataset).
const model = tf.sequential({
layers: [tf.layers.dense({units: 1, inputShape: [10]})]
});
model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
for (i = 1; i < 5 ; ++i) {
const h = await model.fit(tf.ones([8, 10]), tf.ones([8, 1]), {
batchSize: 4,
epochs: 3
});
console.log("Loss after Epoch " + i + " : " + h.history.loss[0]);
}
- x (tf.Tensor|tf.Tensor[]|{[inputName: string]: tf.Tensor}) tf.Tensor of training data, or an array of tf.Tensors if the model has multiple inputs. If all inputs in the model are named, you can also pass a dictionary mapping input names to tf.Tensors.
- y (tf.Tensor|tf.Tensor[]|{[inputName: string]: tf.Tensor}) tf.Tensor of target (label) data, or an array of tf.Tensors if the model has multiple outputs. If all outputs in the model are named, you can also pass a dictionary mapping output names to tf.Tensors.
-
config
(Object)
A
ModelFitConfig
, containing optional fields. Optional - batchSize (number) Number of samples per gradient update. If unspecified, it will default to 32. Optional
- epochs (number) The number of times to iterate over the training data arrays. Optional
- verbose (ModelLoggingVerbosity) Optional
- callbacks (Callback[]|CustomCallbackConfig|CustomCallbackConfig[]) List of callbacks to be called during training. Optional
-
validationSplit
(number)
Float between 0 and 1: fraction of the training data
to be used as validation data. The model will set apart this fraction of
the training data, will not train on it, and will evaluate the loss and
any model metrics on this data at the end of each epoch.
The validation data is selected from the last samples in the
x
andy
data provided, before shuffling. Optional -
validationData
([
tf.Tensor|tf.Tensor[], tf.Tensor|tf.Tensor[]
]|[tf.Tensor | tf.Tensor[], tf.Tensor|tf.Tensor[], tf.Tensor|tf.Tensor[]])
Data on which to evaluate the loss and any model
metrics at the end of each epoch. The model will not be trained on this
data. This could be a tuple [xVal, yVal] or a tuple [xVal, yVal,
valSampleWeights]. The model will not be trained on this data.
validationData
will overridevalidationSplit
. Optional -
shuffle
(boolean)
Whether to shuffle the training data before each epoch. Has
no effect when
stepsPerEpoch
is notnull
. Optional - classWeight ({[classIndex: string]: number}) Optional dictionary mapping class indices (integers) to a weight (float) to apply to the model's loss for the samples from this class during training. This can be useful to tell the model to "pay more attention" to samples from an under-represented class. Optional
- sampleWeight (tf.Tensor) Optional array of the same length as x, containing weights to apply to the model's loss for each sample. In the case of temporal data, you can pass a 2D array with shape (samples, sequenceLength), to apply a different weight to every timestep of every sample. In this case you should make sure to specify sampleWeightMode="temporal" in compile(). Optional
- initialEpoch (number) Epoch at which to start training (useful for resuming a previous training run). Optional
-
stepsPerEpoch
(number)
Total number of steps (batches of samples) before
declaring one epoch finished and starting the next epoch. When training
with Input Tensors such as TensorFlow data tensors, the default
null
is equal to the number of unique samples in your dataset divided by the batch size, or 1 if that cannot be determined. Optional -
validationSteps
(number)
Only relevant if
stepsPerEpoch
is specified. Total number of steps (batches of samples) to validate before stopping. Optional
A model with a stack of layers, feeding linearly from one to the next.
tf.sequential() is a factory function that creates an instance of tf.Sequential.
Adds a layer instance on top of the layer stack.
- layer (tf.layers.Layer) Layer instance.
Returns the loss value & metrics values for the model in test mode.
Loss and metrics are specified during compile()
, which needs to happen
before calls to evaluate()
.
Computation is done in batches.
const model = tf.sequential({
layers: [tf.layers.dense({units: 1, inputShape: [10]})]
});
model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
const result = model.evaluate(tf.ones([8, 10]), tf.ones([8, 1]), {
batchSize: 4,
});
result.print();
-
x
(tf.Tensor|tf.Tensor[])
tf.Tensor of test data, or an
Array
of tf.Tensors if the model has multiple inputs. -
y
(tf.Tensor|tf.Tensor[])
tf.Tensor of target data, or an
Array
of tf.Tensors if the model has multiple outputs. -
config
(Object)
A
ModelEvaluateConfig
, containing optional fields. Optional - batchSize (number) Batch size (Integer). If unspecified, it will default to 32. Optional
- verbose (ModelLoggingVerbosity) Verbosity mode. Optional
- sampleWeight (tf.Tensor) Tensor of weights to weight the contribution of different samples to the loss and metrics. Optional
-
steps
(number)
integer: total number of steps (batches of samples)
before declaring the evaluation round finished. Ignored with the default
value of
undefined
. Optional
Generates output predictions for the input samples.
Computation is done in batches.
Note: the "step" mode of predict() is currently not supported. This is because the TensorFow.js core backend is imperative only.
const model = tf.sequential({
layers: [tf.layers.dense({units: 1, inputShape: [10]})]
});
model.predict(tf.ones([2, 10])).print();
Trains the model for a fixed number of epochs (iterations on a dataset).
const model = tf.sequential({
layers: [tf.layers.dense({units: 1, inputShape: [10]})]
});
model.compile({optimizer: 'sgd', loss: 'meanSquaredError'});
const history = await model.fit(tf.ones([8, 10]), tf.ones([8, 1]), {
batchSize: 4,
epochs: 3
});
- x (tf.Tensor|tf.Tensor[]|{[inputName: string]: tf.Tensor}) tf.Tensor of training data, or an array of tf.Tensors if the model has multiple inputs. If all inputs in the model are named, you can also pass a dictionary mapping input names to tf.Tensors.
- y (tf.Tensor|tf.Tensor[]|{[inputName: string]: tf.Tensor}) tf.Tensor of target (label) data, or an array of tf.Tensors if the model has multiple outputs. If all outputs in the model are named, you can also pass a dictionary mapping output names to tf.Tensors.
-
config
(Object)
A
ModelFitConfig
, containing optional fields. Optional - batchSize (number) Number of samples per gradient update. If unspecified, it will default to 32. Optional
- epochs (number) The number of times to iterate over the training data arrays. Optional
- verbose (ModelLoggingVerbosity) Optional
- callbacks (Callback[]|CustomCallbackConfig|CustomCallbackConfig[]) List of callbacks to be called during training. Optional
-
validationSplit
(number)
Float between 0 and 1: fraction of the training data
to be used as validation data. The model will set apart this fraction of
the training data, will not train on it, and will evaluate the loss and
any model metrics on this data at the end of each epoch.
The validation data is selected from the last samples in the
x
andy
data provided, before shuffling. Optional -
validationData
([
tf.Tensor|tf.Tensor[], tf.Tensor|tf.Tensor[]
]|[tf.Tensor | tf.Tensor[], tf.Tensor|tf.Tensor[], tf.Tensor|tf.Tensor[]])
Data on which to evaluate the loss and any model
metrics at the end of each epoch. The model will not be trained on this
data. This could be a tuple [xVal, yVal] or a tuple [xVal, yVal,
valSampleWeights]. The model will not be trained on this data.
validationData
will overridevalidationSplit
. Optional -
shuffle
(boolean)
Whether to shuffle the training data before each epoch. Has
no effect when
stepsPerEpoch
is notnull
. Optional - classWeight ({[classIndex: string]: number}) Optional dictionary mapping class indices (integers) to a weight (float) to apply to the model's loss for the samples from this class during training. This can be useful to tell the model to "pay more attention" to samples from an under-represented class. Optional
- sampleWeight (tf.Tensor) Optional array of the same length as x, containing weights to apply to the model's loss for each sample. In the case of temporal data, you can pass a 2D array with shape (samples, sequenceLength), to apply a different weight to every timestep of every sample. In this case you should make sure to specify sampleWeightMode="temporal" in compile(). Optional
- initialEpoch (number) Epoch at which to start training (useful for resuming a previous training run). Optional
-
stepsPerEpoch
(number)
Total number of steps (batches of samples) before
declaring one epoch finished and starting the next epoch. When training
with Input Tensors such as TensorFlow data tensors, the default
null
is equal to the number of unique samples in your dataset divided by the batch size, or 1 if that cannot be determined. Optional -
validationSteps
(number)
Only relevant if
stepsPerEpoch
is specified. Total number of steps (batches of samples) to validate before stopping. Optional
tf.SymbolicTensor is a placeholder for a Tensor without any concrete value.
They are most often encountered when building a graph of Layer
s for a
a tf.Model and the input data's shape, but not values are known.
Layers are the primary building block for constructing a Model. Each layer will typically perform some computation to transform its input to its output.
Layers will automatically take care of creating and initializing the various internal variables/weights they need to function.
Applies an activation function to an output.
- config (Object)
- activation ('elu'|'hardsigmoid'|'linear'|'relu'|'relu6'| 'selu'|'sigmoid'|'softmax'|'softplus'|'softsign'|'tanh'|string) Name of the activation function to use.
Creates a dense (fully connected) layer.
This layer implements the operation:
output = activation(dot(input, kernel) + bias)
activation
is the element-wise activation function
passed as the activation
argument.
kernel
is a weights matrix created by the layer.
bias
is a bias vector created by the layer (only applicable if useBias
is true
).
Input shape:
nD tf.Tensor with shape: (batchSize, ..., inputDim)
.
The most common situation would be
a 2D input with shape (batchSize, inputDim)
.
Output shape:
nD tensor with shape: (batchSize, ..., units)
.
For instance, for a 2D input with shape (batchSize, inputDim)
,
the output would have shape (batchSize, units)
.
Note: if the input to the layer has a rank greater than 2, then it is flattened prior to the initial dot product with the kernel.
- config (Object)
- units (number) Positive integer, dimensionality of the output space.
-
activation
('elu'|'hardsigmoid'|'linear'|'relu'|'relu6'|
'selu'|'sigmoid'|'softmax'|'softplus'|'softsign'|'tanh'|string)
Activation function to use.
If unspecified, no activation is applied.
Optional - useBias (boolean) Whether to apply a bias. Optional
- kernelInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the dense kernel weights matrix. Optional
- biasInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the bias vector. Optional
-
inputDim
(number)
If specified, defines inputShape as
[inputDim]
. Optional - kernelConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for the kernel weights. Optional
- biasConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for the bias vector. Optional
- kernelRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the dense kernel weights matrix. Optional
- biasRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the bias vector. Optional
- activityRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the activation. Optional
Applies dropout to the input.
Dropout consists in randomly setting a fraction rate
of input units to 0 at
each update during training time, which helps prevent overfitting.
- config (Object)
- rate (number) Float between 0 and 1. Fraction of the input units to drop.
-
noiseShape
(number[])
Integer array representing the shape of the binary dropout mask that will
be multiplied with the input.
For instance, if your inputs have shape
Optional(batchSize, timesteps, features)
and you want the dropout mask to be the same for all timesteps, you can usenoise_shape=(batch_size, 1, features)
. - seed (number) An integer to use as random seed. Optional
Maps positive integers (indices) into dense vectors of fixed size. eg. [[4], [20]] -> [[0.25, 0.1], [0.6, -0.2]]
Input shape: 2D tensor with shape: [batchSize, sequenceLength]
.
Output shape: 3D tensor with shape: [batchSize, sequenceLength, outputDim]
.
- config (Object)
- inputDim (number) Integer > 0. Size of the vocabulary, i.e. maximum integer index + 1.
- outputDim (number) Integer >= 0. Dimension of the dense embedding.
-
embeddingsInitializer
('constant'|'glorotNormal'|'glorotUniform'|
'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'|
'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer)
Initializer for the
embeddings
matrix. Optional -
embeddingsRegularizer
('l1l2'|string|Regularizer)
Regularizer function applied to the
embeddings
matrix. Optional - activityRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the activation. Optional
-
embeddingsConstraint
('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint)
Constraint function applied to the
embeddings
matrix. Optional -
maskZero
(boolean)
Whether the input value 0 is a special "padding" value that should be
masked out. This is useful when using recurrent layers which may take
variable length input.
If this is
OptionalTrue
then all subsequent layers in the model need to support masking or an exception will be raised. If maskZero is set toTrue
, as a consequence, index 0 cannot be used in the vocabulary (inputDim should equal size of vocabulary + 1). -
inputLength
(number|number[])
Length of input sequences, when it is constant.
This argument is required if you are going to connect
Optionalflatten
thendense
layers upstream (without it, the shape of the dense outputs cannot be computed).
Flattens the input. Does not affect the batch size.
A Flatten
layer flattens each batch in its inputs to 1D (making the output
2D).
For example:
const input = tf.input({shape: [4, 3]});
const flattenLayer = tf.layers.flatten();
// Inspect the inferred output shape of the flatten layer, which
// equals `[null, 12]`. The 2nd dimension is 4 * 3, i.e., the result of the
// flattening. (The 1st dimension is the undermined batch size.)
console.log(flattenLayer.apply(input).shape);
- config (Object) Optional
-
inputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchInputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchSize
(number)
If
inputShape
is specified andbatchInputShape
is not specifiedd,batchSize
is used to construct thebatchInputShape
:[batchSize, ...inputShape]
Optional - dtype ('float32'|'int32'|'bool') The data-type for this layer. Defaults to 'float32'. This argument is only applicable to input layers (the first layer of a model). Optional
- name (string) Name for this layer. Optional
- trainable (boolean) Whether this layer is trainable. Defaults to true. Optional
-
updatable
(boolean)
Whether the weights of this layer are updatable by
fit
. Optional - weights (tf.Tensor[]) Initial weight values of the layer. Optional
- inputDType ('float32'|'int32'|'bool') Legacy support. Do not use for new code. Optional
Repeat the input n times.
- config (Object)
- n (number) The integer number of times to repeat the input.
1D convolution layer (e.g., temporal convolution).
This layer creates a convolution kernel that is convolved with the layer input over a single spatial (or temporal) dimension to produce a tensor of outputs.
If use_bias
is True, a bias vector is created and added to the outputs.
If activation
is not null
, it is applied to the outputs as well.
When using this layer as the first layer in a model, provide an inputShape
argument Array
or null
.
For example, inputShape
would be:
[10, 128]
for sequences of 10 vectors of 128-dimensional vectors[null, 128]
for variable-length sequences of 128-dimensional vectors.
- config (Object)
- kernelSize (number|number[]) The dimensions of the convolution window. If kernelSize is a number, the convolutional window will be square.
- filters (number) The dimensionality of the output space (i.e. the number of filters in the convolution). Optional
-
strides
(number|number[])
The strides of the convolution in each dimension. If strides is a number,
strides in both dimensions are equal.
Specifying any stride value != 1 is incompatible with specifying any
OptionaldilationRate
value != 1. - padding ('valid'|'same'|'casual') Padding mode. Optional
-
dataFormat
('channelsFirst'|'channelsLast')
Format of the data, which determines the ordering of the dimensions in
the inputs.
channels_last
corresponds to inputs with shape(batch, ..., channels)
channels_first
corresponds to inputs with shape(batch, channels, ...)
.Defaults to
Optionalchannels_last
. -
dilationRate
(number|number[])
The dilation rate to use for the dilated convolution in each dimension.
Should be an integer or array of integers.
Currently, specifying any
OptionaldilationRate
value != 1 is incompatible with specifying anystrides
value != 1. -
activation
(string)
Activation function of the layer.
If you don't specify the activation, none is applied.
Optional - useBias (boolean) Whether the layer uses a bias vector. Defaults to false. Optional
- kernelInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the convolutional kernel weights matrix. Optional
- biasInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the bias vector. Optional
- kernelConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for the convolutional kernel weights. Optional
- biasConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for the bias vector. Optional
- kernelRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the kernel weights matrix. Optional
- biasRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the bias vector. Optional
- activityRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the activation. Optional
2D convolution layer (e.g. spatial convolution over images).
This layer creates a convolution kernel that is convolved with the layer input to produce a tensor of outputs.
If useBias
is True, a bias vector is created and added to the outputs.
If activation
is not null
, it is applied to the outputs as well.
When using this layer as the first layer in a model,
provide the keyword argument inputShape
(Array of integers, does not include the sample axis),
e.g. inputShape=[128, 128, 3]
for 128x128 RGB pictures
in dataFormat='channelsLast'
.
- config (Object)
- kernelSize (number|number[]) The dimensions of the convolution window. If kernelSize is a number, the convolutional window will be square.
- filters (number) The dimensionality of the output space (i.e. the number of filters in the convolution). Optional
-
strides
(number|number[])
The strides of the convolution in each dimension. If strides is a number,
strides in both dimensions are equal.
Specifying any stride value != 1 is incompatible with specifying any
OptionaldilationRate
value != 1. - padding ('valid'|'same'|'casual') Padding mode. Optional
-
dataFormat
('channelsFirst'|'channelsLast')
Format of the data, which determines the ordering of the dimensions in
the inputs.
channels_last
corresponds to inputs with shape(batch, ..., channels)
channels_first
corresponds to inputs with shape(batch, channels, ...)
.Defaults to
Optionalchannels_last
. -
dilationRate
(number|number[])
The dilation rate to use for the dilated convolution in each dimension.
Should be an integer or array of integers.
Currently, specifying any
OptionaldilationRate
value != 1 is incompatible with specifying anystrides
value != 1. -
activation
(string)
Activation function of the layer.
If you don't specify the activation, none is applied.
Optional - useBias (boolean) Whether the layer uses a bias vector. Defaults to false. Optional
- kernelInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the convolutional kernel weights matrix. Optional
- biasInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the bias vector. Optional
- kernelConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for the convolutional kernel weights. Optional
- biasConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for the bias vector. Optional
- kernelRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the kernel weights matrix. Optional
- biasRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the bias vector. Optional
- activityRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the activation. Optional
Depthwise separable 2D convolution.
Depthwise Separable convolutions consists in performing just the first step
in a depthwise spatial convolution (which acts on each input channel
separately). The depthMultplier
argument controls how many output channels
are generated per input channel in the depthwise step.
- config (Object)
- kernelSize (number|[number, number]) An integer or Array of 2 integers, specifying the width and height of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions.
-
depthMultiplier
(number)
The number of depthwise convolution output channels for each input
channel.
The total number of depthwise convolution output channels will be equal to
filtersIn * depthMultiplier
. Default: 1. Optional - depthwiseInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the depthwise kernel matrix. Default: GlorotNormal. Optional
- depthwiseConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for the depthwise kernel matrix. Optional
- depthwiseRegularizer ('l1l2'|string|Regularizer) Regulzarizer function for the depthwise kernel matrix. Optional
Layer that performs element-wise addition on an Array
of inputs.
It takes as input a list of tensors, all of the same shape, and returns a
single tensor (also of the same shape). The inputs are specified as an
Array
when the apply
method of the Add
layer instance is called. For
example:
const input1 = tf.input({shape: [2, 2]});
const input2 = tf.input({shape: [2, 2]});
const addLayer = tf.layers.add();
const sum = addLayer.apply([input1, input2]);
console.log(sum.shape);
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.
- config (Object) Optional
-
inputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchInputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchSize
(number)
If
inputShape
is specified andbatchInputShape
is not specifiedd,batchSize
is used to construct thebatchInputShape
:[batchSize, ...inputShape]
Optional - dtype ('float32'|'int32'|'bool') The data-type for this layer. Defaults to 'float32'. This argument is only applicable to input layers (the first layer of a model). Optional
- name (string) Name for this layer. Optional
- trainable (boolean) Whether this layer is trainable. Defaults to true. Optional
-
updatable
(boolean)
Whether the weights of this layer are updatable by
fit
. Optional - weights (tf.Tensor[]) Initial weight values of the layer. Optional
- inputDType ('float32'|'int32'|'bool') Legacy support. Do not use for new code. Optional
Layer that performs element-wise averaging on an Array
of inputs.
It takes as input a list of tensors, all of the same shape, and returns a single tensor (also of the same shape). For example:
const input1 = tf.input({shape: [2, 2]});
const input2 = tf.input({shape: [2, 2]});
const averageLayer = tf.layers.average();
const average = averageLayer.apply([input1, input2]);
console.log(average.shape);
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.
- config (Object) Optional
-
inputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchInputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchSize
(number)
If
inputShape
is specified andbatchInputShape
is not specifiedd,batchSize
is used to construct thebatchInputShape
:[batchSize, ...inputShape]
Optional - dtype ('float32'|'int32'|'bool') The data-type for this layer. Defaults to 'float32'. This argument is only applicable to input layers (the first layer of a model). Optional
- name (string) Name for this layer. Optional
- trainable (boolean) Whether this layer is trainable. Defaults to true. Optional
-
updatable
(boolean)
Whether the weights of this layer are updatable by
fit
. Optional - weights (tf.Tensor[]) Initial weight values of the layer. Optional
- inputDType ('float32'|'int32'|'bool') Legacy support. Do not use for new code. Optional
Layer that concatenates an Array
of inputs.
It takes a list of tensors, all of the same shape except for the concatenation axis, and returns a single tensor, the concatenation of all inputs. For example:
const input1 = tf.input({shape: [2, 2]});
const input2 = tf.input({shape: [2, 3]});
const concatLayer = tf.layers.concatenate();
const output = concatLayer.apply([input1, input2]);
console.log(output.shape);
// You get [null, 2, 5], with the first dimension as the undetermined batch
// dimension. The last dimension (5) is the result of concatenating the
// last dimensions of the inputs (2 and 3).
- config (Object)
- axis (number) Axis along which to concatenate. Optional
Layer that computes the element-wise maximum an Array
of inputs.
It takes as input a list of tensors, all of the same shape and returns a single tensor (also of the same shape). For example:
const input1 = tf.input({shape: [2, 2]});
const input2 = tf.input({shape: [2, 2]});
const maxLayer = tf.layers.maximum();
const max = maxLayer.apply([input1, input2]);
console.log(max.shape);
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.
- config (Object) Optional
-
inputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchInputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchSize
(number)
If
inputShape
is specified andbatchInputShape
is not specifiedd,batchSize
is used to construct thebatchInputShape
:[batchSize, ...inputShape]
Optional - dtype ('float32'|'int32'|'bool') The data-type for this layer. Defaults to 'float32'. This argument is only applicable to input layers (the first layer of a model). Optional
- name (string) Name for this layer. Optional
- trainable (boolean) Whether this layer is trainable. Defaults to true. Optional
-
updatable
(boolean)
Whether the weights of this layer are updatable by
fit
. Optional - weights (tf.Tensor[]) Initial weight values of the layer. Optional
- inputDType ('float32'|'int32'|'bool') Legacy support. Do not use for new code. Optional
Layer that computes the element-wise minimum of an Array
of inputs.
It takes as input a list of tensors, all of the same shape and returns a single tensor (also of the same shape). For example:
const input1 = tf.input({shape: [2, 2]});
const input2 = tf.input({shape: [2, 2]});
const minLayer = tf.layers.minimum();
const min = minLayer.apply([input1, input2]);
console.log(min.shape);
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.
- config (Object) Optional
-
inputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchInputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchSize
(number)
If
inputShape
is specified andbatchInputShape
is not specifiedd,batchSize
is used to construct thebatchInputShape
:[batchSize, ...inputShape]
Optional - dtype ('float32'|'int32'|'bool') The data-type for this layer. Defaults to 'float32'. This argument is only applicable to input layers (the first layer of a model). Optional
- name (string) Name for this layer. Optional
- trainable (boolean) Whether this layer is trainable. Defaults to true. Optional
-
updatable
(boolean)
Whether the weights of this layer are updatable by
fit
. Optional - weights (tf.Tensor[]) Initial weight values of the layer. Optional
- inputDType ('float32'|'int32'|'bool') Legacy support. Do not use for new code. Optional
Layer that multiplies (element-wise) an Array
of inputs.
It takes as input an Array of tensors, all of the same shape, and returns a single tensor (also of the same shape). For example:
const input1 = tf.input({shape: [2, 2]});
const input2 = tf.input({shape: [2, 2]});
const input3 = tf.input({shape: [2, 2]});
const multiplyLayer = tf.layers.multiply();
const product = multiplyLayer.apply([input1, input2, input3]);
console.log(product.shape);
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.
- config (Object) Optional
-
inputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchInputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchSize
(number)
If
inputShape
is specified andbatchInputShape
is not specifiedd,batchSize
is used to construct thebatchInputShape
:[batchSize, ...inputShape]
Optional - dtype ('float32'|'int32'|'bool') The data-type for this layer. Defaults to 'float32'. This argument is only applicable to input layers (the first layer of a model). Optional
- name (string) Name for this layer. Optional
- trainable (boolean) Whether this layer is trainable. Defaults to true. Optional
-
updatable
(boolean)
Whether the weights of this layer are updatable by
fit
. Optional - weights (tf.Tensor[]) Initial weight values of the layer. Optional
- inputDType ('float32'|'int32'|'bool') Legacy support. Do not use for new code. Optional
Batch normalization layer (Ioffe and Szegedy, 2014).
Normalize the activations of the previous layer at each batch, i.e. applies a transformation that maintains the mean activation close to 0 and the activation standard deviation close to 1.
Input shape:
Arbitrary. Use the keyword argument inputShape
(Array of integers, does
not include the sample axis) when calling the constructor of this class,
if this layer is used as a first layer in a model.
Output shape: Same shape as input.
References:
- config (Object)
-
axis
(number)
The integer axis that should be normalized (typically the features axis).
Defaults to -1.
For instance, after a
OptionalConv2D
layer withdata_format="channels_first"
, setaxis=1
in tf.batchNormalization(). - momentum (number) Momentum of the moving average. Defaults to 0.99. Optional
- epsilon (number) Small float added to the variance to avoid dividing by zero. Defaults to 1e-3. Optional
-
center
(boolean)
If
true
, add offset ofbeta
to normalized tensor. Iffalse
,beta
is ignored. Defaults to true. Optional -
scale
(boolean)
If
true
, multiply bygamma
. Iffalse
,gamma
is not used. When the next layer is linear (also e.g.nn.relu
), this can be disabled since the scaling will be done by the next layer. Defaults to true. Optional - betaInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the beta weight. Defaults to 'zeros'. Optional
- gammaInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the gamma weight. Defaults to tf.ones(). Optional
- movingMeanInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the moving mean. Defaults to tf.zeros() Optional
- movingVarianceInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the moving variance. Defaults to 'Ones'. Optional
- betaConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for the beta weight. Optional
- gammaConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint for gamma weight. Optional
- betaRegularizer ('l1l2'|string|Regularizer) Regularizer for the beta weight. Optional
- gammaRegularizer ('l1l2'|string|Regularizer) Regularizer for the gamma weight. Optional
Average pooling operation for spatial data.
Input shape: [batchSize, inLength, channels]
Output shape: [batchSize, pooledLength, channels]
- config (Object)
- poolSize (number) Size of the window to pool over, should be an integer. Optional
-
strides
(number)
Period at which to sample the pooled values.
If
Optionalnull
, defaults topoolSize
. - padding ('valid'|'same'|'casual') How to fill in data that's not an integer multiple of poolSize. Optional
Average pooling operation for spatial data.
Input shape:
- If
dataFormat === CHANNEL_LAST
: 4D tensor with shape:[batchSize, rows, cols, channels]
- If
dataFormat === CHANNEL_FIRST
: 4D tensor with shape:[batchSize, channels, rows, cols]
Output shape
- If
dataFormat === CHANNEL_LAST
: 4D tensor with shape:[batchSize, pooleRows, pooledCols, channels]
- If
dataFormat === CHANNEL_FIRST
: 4D tensor with shape:[batchSize, channels, pooleRows, pooledCols]
- config (Object)
-
poolSize
(number|[number, number])
Factors by which to downscale in each dimension [vertical, horizontal].
Expects an integer or an array of 2 integers.
For example,
Optional[2, 2]
will halve the input in both spatial dimension. If only one integer is specified, the same window length will be used for both dimensions. -
strides
([number, number])
The size of the stride in each dimension of the pooling window. Expects an
integer or an array of 2 integers. Integer, tuple of 2 integers, or None.
If
Optionalnull
, defaults topoolSize
. - padding ('valid'|'same'|'casual') The padding type to use for the pooling layer. Optional
- dataFormat ('channelsFirst'|'channelsLast') The data format to use for the pooling layer. Optional
Global average pooling operation for temporal data.
Input Shape: 3D tensor with shape: [batchSize, steps, features]
.
Output Shape:2D tensor with shape: [batchSize, features]
.
- config (Object)
-
inputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchInputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchSize
(number)
If
inputShape
is specified andbatchInputShape
is not specifiedd,batchSize
is used to construct thebatchInputShape
:[batchSize, ...inputShape]
Optional - dtype ('float32'|'int32'|'bool') The data-type for this layer. Defaults to 'float32'. This argument is only applicable to input layers (the first layer of a model). Optional
- name (string) Name for this layer. Optional
- trainable (boolean) Whether this layer is trainable. Defaults to true. Optional
-
updatable
(boolean)
Whether the weights of this layer are updatable by
fit
. Optional - weights (tf.Tensor[]) Initial weight values of the layer. Optional
- inputDType ('float32'|'int32'|'bool') Legacy support. Do not use for new code. Optional
Global average pooling operation for spatial data.
Input shape:
- If
dataFormat
isCHANNEL_LAST
: 4D tensor with shape:[batchSize, rows, cols, channels]
. - If
dataFormat
isCHANNEL_FIRST
: 4D tensor with shape:[batchSize, channels, rows, cols]
.
Output shape:
2D tensor with shape: [batchSize, channels]
.
- config (Object)
-
dataFormat
('channelsFirst'|'channelsLast')
One of
CHANNEL_LAST
(default) orCHANNEL_FIRST
.The ordering of the dimensions in the inputs.
OptionalCHANNEL_LAST
corresponds to inputs with shape[batch, height, width, channels[
whileCHANNEL_FIRST
corresponds to inputs with shape[batch, channels, height, width]
.
Global max pooling operation for temporal data.
Input Shape: 3D tensor with shape: [batchSize, steps, features]
.
Output Shape:2D tensor with shape: [batchSize, features]
.
- config (Object)
-
inputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchInputShape
(number[])
If defined, will be used to create an input layer to insert before this
layer. If both
inputShape
andbatchInputShape
are defined,batchInputShape
will be used. This argument is only applicable to input layers (the first layer of a model). Optional -
batchSize
(number)
If
inputShape
is specified andbatchInputShape
is not specifiedd,batchSize
is used to construct thebatchInputShape
:[batchSize, ...inputShape]
Optional - dtype ('float32'|'int32'|'bool') The data-type for this layer. Defaults to 'float32'. This argument is only applicable to input layers (the first layer of a model). Optional
- name (string) Name for this layer. Optional
- trainable (boolean) Whether this layer is trainable. Defaults to true. Optional
-
updatable
(boolean)
Whether the weights of this layer are updatable by
fit
. Optional - weights (tf.Tensor[]) Initial weight values of the layer. Optional
- inputDType ('float32'|'int32'|'bool') Legacy support. Do not use for new code. Optional
Global max pooling operation for spatial data.
Input shape:
- If
dataFormat
isCHANNEL_LAST
: 4D tensor with shape:[batchSize, rows, cols, channels]
. - If
dataFormat
isCHANNEL_FIRST
: 4D tensor with shape:[batchSize, channels, rows, cols]
.
Output shape:
2D tensor with shape: [batchSize, channels]
.
- config (Object)
-
dataFormat
('channelsFirst'|'channelsLast')
One of
CHANNEL_LAST
(default) orCHANNEL_FIRST
.The ordering of the dimensions in the inputs.
OptionalCHANNEL_LAST
corresponds to inputs with shape[batch, height, width, channels[
whileCHANNEL_FIRST
corresponds to inputs with shape[batch, channels, height, width]
.
Max pooling operation for temporal data.
Input shape: [batchSize, inLength, channels]
Output shape: [batchSize, pooledLength, channels]
- config (Object)
- poolSize (number) Size of the window to pool over, should be an integer. Optional
-
strides
(number)
Period at which to sample the pooled values.
If
Optionalnull
, defaults topoolSize
. - padding ('valid'|'same'|'casual') How to fill in data that's not an integer multiple of poolSize. Optional
Max pooling operation for spatial data.
Input shape
- If
dataFormat === CHANNEL_LAST
: 4D tensor with shape:[batchSize, rows, cols, channels]
- If
dataFormat === CHANNEL_FIRST
: 4D tensor with shape:[batchSize, channels, rows, cols]
Output shape
- If
dataFormat=CHANNEL_LAST
: 4D tensor with shape:[batchSize, pooleRows, pooledCols, channels]
- If
dataFormat=CHANNEL_FIRST
: 4D tensor with shape:[batchSize, channels, pooleRows, pooledCols]
- config (Object)
-
poolSize
(number|[number, number])
Factors by which to downscale in each dimension [vertical, horizontal].
Expects an integer or an array of 2 integers.
For example,
Optional[2, 2]
will halve the input in both spatial dimension. If only one integer is specified, the same window length will be used for both dimensions. -
strides
([number, number])
The size of the stride in each dimension of the pooling window. Expects an
integer or an array of 2 integers. Integer, tuple of 2 integers, or None.
If
Optionalnull
, defaults topoolSize
. - padding ('valid'|'same'|'casual') The padding type to use for the pooling layer. Optional
- dataFormat ('channelsFirst'|'channelsLast') The data format to use for the pooling layer. Optional
Gated Recurrent Unit - Cho et al. 2014.
This is an RNN
layer consisting of one GRUCell
. However, unlike
the underlying GRUCell
, the apply
method of SimpleRNN
operates
on a sequence of inputs. The shape of the input (not including the first,
batch dimension) needs to be at least 2-D, with the first dimension being
time steps. For example:
const rnn = tf.layers.gru({units: 8, returnSequences: true});
// Create an input with 10 time steps.
const input = tf.input({shape: [10, 20]});
const output = rnn.apply(input);
console.log(output);
// [null, 10, 8]: 1st dimension is unknown batch size; 2nd dimension is the
// same as the sequence length of [tf.input()](#input), due to `returnSequences`: `true`;
// 3rd dimension is the `GRUCell`'s number of units.
- config (Object)
-
implementation
(number)
Implementation mode, either 1 or 2.
Mode 1 will structure its operations as a larger number of smaller dot products and additions.
Mode 2 will batch them into fewer, larger operations. These modes will have different performance profiles on different hardware and for different applications.
Optional
Cell class for GRU
.
GRUCell
is distinct from the RNN
subclass GRU
in that its
apply
method takes the input data of only a single time step and returns
the cell's output at the time step, while GRU
takes the input data
over a number of time steps. For example:
const cell = tf.layers.gruCell({units: 2});
const input = tf.input({shape: [10]});
const output = cell.apply(input);
console.log(output.shape);
// [null, 10]: This is the cell's output at a single time step. The 1st
// dimension is the unknown batch size.
Instance(s) of GRUCell
can be used to construct RNN
layers. The
most typical use of this workflow is to combine a number of cells into a
stacked RNN cell (i.e., StackedRNNCell
internally) and use it to create an
RNN. For example:
const cells = [
tf.layers.gruCell({units: 4}),
tf.layers.gruCell({units: 8}),
];
const rnn = tf.layers.rnn({cell: cells, returnSequences: true});
// Create an input with 10 time steps and a length-20 vector at each step.
const input = tf.input({shape: [10, 20]});
const output = rnn.apply(input);
console.log(output);
// [null, 10, 8]: 1st dimension is unknown batch size; 2nd dimension is the
// same as the sequence length of [tf.input()](#input), due to `returnSequences`: `true`;
// 3rd dimension is the last `gruCell`'s number of units.
To create an RNN
consisting of only one GRUCell
, use the
tf.layers.gru
.
- config (Object)
-
recurrentActivation
(string)
Activation function to use for the recurrent step.
Defaults to hard sigmoid (
hardSigomid
).If
Optionalnull
, no activation is applied. -
implementation
(number)
Implementation mode, either 1 or 2.
Mode 1 will structure its operations as a larger number of smaller dot products and additions.
Mode 2 will batch them into fewer, larger operations. These modes will have different performance profiles on different hardware and for different applications.
Optional
Long-Short Term Memory layer - Hochreiter 1997.
This is an RNN
layer consisting of one LSTMCell
. However, unlike
the underlying LSTMCell
, the apply
method of LSTM
operates
on a sequence of inputs. The shape of the input (not including the first,
batch dimension) needs to be at least 2-D, with the first dimension being
time steps. For example:
const lstm = tf.layers.lstm({units: 8, returnSequences: true});
// Create an input with 10 time steps.
const input = tf.input({shape: [10, 20]});
const output = rnn.apply(input);
console.log(output);
// [null, 10, 8]: 1st dimension is unknown batch size; 2nd dimension is the
// same as the sequence length of [tf.input()](#input), due to `returnSequences`: `true`;
// 3rd dimension is the `LSTMCell`'s number of units.
- config (Object)
-
unitForgetBias
(boolean)
If
true
, add 1 to the bias of the forget gate at initialization. Setting it totrue
will also forcebiasInitializer = 'zeros'
. This is recommended in Jozefowicz et al.. Optional - implementation (1|2) Implementation mode, either 1 or 2. Mode 1 will structure its operations as a larger number of smaller dot products and additions, whereas mode 2 will batch them into fewer, larger operations. These modes will have different performance profiles on different hardware and for different applications. Optional
Cell class for LSTM
.
LSTMCell
is distinct from the RNN
subclass LSTM
in that its
apply
method takes the input data of only a single time step and returns
the cell's output at the time step, while LSTM
takes the input data
over a number of time steps. For example:
const cell = tf.layers.lstmCell({units: 2});
const input = tf.input({shape: [10]});
const output = cell.apply(input);
console.log(output.shape);
// [null, 10]: This is the cell's output at a single time step. The 1st
// dimension is the unknown batch size.
Instance(s) of LSTMCell
can be used to construct RNN
layers. The
most typical use of this workflow is to combine a number of cells into a
stacked RNN cell (i.e., StackedRNNCell
internally) and use it to create an
RNN. For example:
const cells = [
tf.layers.lstmCell({units: 4}),
tf.layers.lstmCell({units: 8}),
];
const rnn = tf.layers.rnn({cell: cells, returnSequences: true});
// Create an input with 10 time steps and a length-20 vector at each step.
const input = tf.input({shape: [10, 20]});
const output = rnn.apply(input);
console.log(output);
// [null, 10, 8]: 1st dimension is unknown batch size; 2nd dimension is the
// same as the sequence length of [tf.input()](#input), due to `returnSequences`: `true`;
// 3rd dimension is the last `lstmCell`'s number of units.
To create an RNN
consisting of only one LSTMCell
, use the
tf.layers.lstm
.
- config (Object)
-
recurrentActivation
('elu'|'hardsigmoid'|'linear'|'relu'|'relu6'|
'selu'|'sigmoid'|'softmax'|'softplus'|'softsign'|'tanh'|string)
Activation function to use for the recurrent step.
Defaults to hard sigmoid (
hardSigomid
).If
Optionalnull
, no activation is applied. -
unitForgetBias
(boolean)
If
true
, add 1 to the bias of the forget gate at initialization. Setting it totrue
will also forcebiasInitializer = 'zeros'
. This is recommended in Jozefowicz et al.. Optional -
implementation
(1|2)
Implementation mode, either 1 or 2.
Mode 1 will structure its operations as a larger number of smaller dot products and additions.
Mode 2 will batch them into fewer, larger operations. These modes will have different performance profiles on different hardware and for different applications.
Optional
Base class for recurrent layers.
Input shape:
3D tensor with shape [batchSize, timeSteps, inputDim]
.
Output shape:
- if
returnState
, an Array of tensors (i.e., tf.Tensors). The first tensor is the output. The remaining tensors are the states at the last time step, each with shape[batchSize, units]
. - if
returnSequences
, the output will have shape[batchSize, timeSteps, units]
. - else, the output will have shape
[batchSize, units]
.
Masking:
This layer supports masking for input data with a variable number
of timesteps. To introduce masks to your data,
use an embedding layer with the mask_zero
parameter
set to True
.
Notes on using statefulness in RNNs: You can set RNN layers to be 'stateful', which means that the states computed for the samples in one batch will be reused as initial states for the samples in the next batch. This assumes a one-to-one mapping between samples in different successive batches.
To enable statefulness:
- specify stateful: true
in the layer constructor.
- specify a fixed batch size for your model, by passing
if sequential model:
batchInputShape=[...]
to the first layer in your model.
else for functional model with 1 or more Input layers:
batchShape=[...]
to all the first layers in your model.
This is the expected shape of your inputs including the batch size.
It should be a tuple of integers, e.g. (32, 10, 100)
.
- specify shuffle=False
when calling fit().
To reset the states of your model, call .reset_states()
on either
a specific layer, or on your entire model.
Note on specifying the initial state of RNNs
You can specify the initial state of RNN layers symbolically by
calling them with the option initialState
. The value of
initialState
should be a tensor or list of tensors representing
the initial state of the RNN layer.
You can specify the initial state of RNN layers numerically by
calling resetStates
with the keyword argument states
. The value of
states
should be a numpy array or list of numpy arrays representing
the initial state of the RNN layer.
Note on passing external constants to RNNs
You can pass "external" constants to the cell using the constants
keyword argument of RNN.call
method. This requires that the cell.call
method accepts the same keyword argument constants
. Such constants
can be used to conditon the cell transformation on additional static inputs
(not changing over time), a.k.a an attention mechanism.
- config (Object)
- cell (tf.RNNCell|tf.RNNCell[])
Fully-connected RNN where the output is to be fed back to input.
This is an RNN
layer consisting of one SimpleRNNCell
. However, unlike
the underlying SimpleRNNCell
, the apply
method of SimpleRNN
operates
on a sequence of inputs. The shape of the input (not including the first,
batch dimension) needs to be at least 2-D, with the first dimension being
time steps. For example:
const rnn = tf.layers.simpleRNN({units: 8, returnSequences: true});
// Create an input with 10 time steps.
const input = tf.input({shape: [10, 20]});
const output = rnn.apply(input);
console.log(output);
// [null, 10, 8]: 1st dimension is unknown batch size; 2nd dimension is the
// same as the sequence length of [tf.input()](#input), due to `returnSequences`: `true`;
// 3rd dimension is the `SimpleRNNCell`'s number of units.
- config (Object)
- units (number) Positive integer, dimensionality of the output space.
-
activation
('elu'|'hardsigmoid'|'linear'|'relu'|'relu6'|
'selu'|'sigmoid'|'softmax'|'softplus'|'softsign'|'tanh'|string)
Activation function to use.
Defaults to hyperbolic tangent (tf.tanh())
If you pass
Optionalnull
, no activation will be applied. - useBias (boolean) Whether the layer uses a bias vector. Optional
-
kernelInitializer
('constant'|'glorotNormal'|'glorotUniform'|
'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'|
'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer)
Initializer for the
kernel
weights matrix, used for the linear transformation of the inputs. Optional -
recurrentInitializer
('constant'|'glorotNormal'|'glorotUniform'|
'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'|
'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer)
Initializer for the
recurrentKernel
weights matrix, used for linear transformation of the recurrent state. Optional - biasInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the bias vector. Optional
- kernelRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the kernel weights matrix. Optional
- recurrentRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the recurrentKernel weights matrix. Optional
- biasRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the bias vector. Optional
- kernelConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint function applied to the kernel weights matrix. Optional
- recurrentConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint function applied to the recurrentKernel weights matrix. Optional
- biasConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraint function applied to the bias vector. Optional
- dropout (number) Number between 0 and 1. Fraction of the units to drop for the linear transformation of the inputs. Optional
- recurrentDropout (number) Number between 0 and 1. Fraction of the units to drop for the linear transformation of the recurrent state. Optional
Cell class for SimpleRNN
.
SimpleRNNCell
is distinct from the RNN
subclass SimpleRNN
in that its
apply
method takes the input data of only a single time step and returns
the cell's output at the time step, while SimpleRNN
takes the input data
over a number of time steps. For example:
const cell = tf.layers.simpleRNNCell({units: 2});
const input = tf.input({shape: [10]});
const output = cell.apply(input);
console.log(output.shape);
// [null, 10]: This is the cell's output at a single time step. The 1st
// dimension is the unknown batch size.
Instance(s) of SimpleRNNCell
can be used to construct RNN
layers. The
most typical use of this workflow is to combine a number of cells into a
stacked RNN cell (i.e., StackedRNNCell
internally) and use it to create an
RNN. For example:
const cells = [
tf.layers.simpleRNNCell({units: 4}),
tf.layers.simpleRNNCell({units: 8}),
];
const rnn = tf.layers.rnn({cell: cells, returnSequences: true});
// Create an input with 10 time steps and a length-20 vector at each step.
const input = tf.input({shape: [10, 20]});
const output = rnn.apply(input);
console.log(output);
// [null, 10, 8]: 1st dimension is unknown batch size; 2nd dimension is the
// same as the sequence length of [tf.input()](#input), due to `returnSequences`: `true`;
// 3rd dimension is the last `SimpleRNNCell`'s number of units.
To create an RNN
consisting of only one SimpleRNNCell
, use the
tf.layers.simpleRNN
.
- config (Object)
- units (number) units: Positive integer, dimensionality of the output space.
-
activation
('elu'|'hardsigmoid'|'linear'|'relu'|'relu6'|
'selu'|'sigmoid'|'softmax'|'softplus'|'softsign'|'tanh'|string)
Activation function to use.
Default: hyperbolic tangent ('tanh').
If you pass
null
, 'linear' activation will be applied. Optional - useBias (boolean) Whether the layer uses a bias vector. Optional
-
kernelInitializer
('constant'|'glorotNormal'|'glorotUniform'|
'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'|
'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer)
Initializer for the
kernel
weights matrix, used for the linear transformation of the inputs. Optional -
recurrentInitializer
('constant'|'glorotNormal'|'glorotUniform'|
'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'|
'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer)
Initializer for the
recurrentKernel
weights matrix, used for linear transformation of the recurrent state. Optional - biasInitializer ('constant'|'glorotNormal'|'glorotUniform'| 'heNormal'|'identity'|'leCunNormal'|'ones'|'orthogonal'|'randomNormal'| 'randomUniform'|'truncatedNormal'|'varianceScaling'|'zeros'|string|tf.initializers.Initializer) Initializer for the bias vector. Optional
-
kernelRegularizer
('l1l2'|string|Regularizer)
Regularizer function applied to the
kernel
weights matrix. Optional -
recurrentRegularizer
('l1l2'|string|Regularizer)
Regularizer function applied to the
recurrent_kernel
weights matrix. Optional - biasRegularizer ('l1l2'|string|Regularizer) Regularizer function applied to the bias vector. Optional
-
kernelConstraint
('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint)
Constraint function applied to the
kernel
weights matrix. Optional -
recurrentConstraint
('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint)
Constraint function applied to the
recurrentKernel
weights matrix. Optional - biasConstraint ('maxNorm'|'minMaxNorm'|'nonNeg'|'unitNorm'|string|tf.constraints.Constraint) Constraintfunction applied to the bias vector. Optional
- dropout (number) Float number between 0 and 1. Fraction of the units to drop for the linear transformation of the inputs. Optional
- recurrentDropout (number) Float number between 0 and 1. Fraction of the units to drop for the linear transformation of the recurrent state. Optional
Base class for recurrent layers.
Input shape:
3D tensor with shape [batchSize, timeSteps, inputDim]
.
Output shape:
- if
returnState
, an Array of tensors (i.e., tf.Tensors). The first tensor is the output. The remaining tensors are the states at the last time step, each with shape[batchSize, units]
. - if
returnSequences
, the output will have shape[batchSize, timeSteps, units]
. - else, the output will have shape
[batchSize, units]
.
Masking:
This layer supports masking for input data with a variable number
of timesteps. To introduce masks to your data,
use an embedding layer with the mask_zero
parameter
set to True
.
Notes on using statefulness in RNNs: You can set RNN layers to be 'stateful', which means that the states computed for the samples in one batch will be reused as initial states for the samples in the next batch. This assumes a one-to-one mapping between samples in different successive batches.
To enable statefulness:
- specify stateful: true
in the layer constructor.
- specify a fixed batch size for your model, by passing
if sequential model:
batchInputShape=[...]
to the first layer in your model.
else for functional model with 1 or more Input layers:
batchShape=[...]
to all the first layers in your model.
This is the expected shape of your inputs including the batch size.
It should be a tuple of integers, e.g. (32, 10, 100)
.
- specify shuffle=False
when calling fit().
To reset the states of your model, call .reset_states()
on either
a specific layer, or on your entire model.
Note on specifying the initial state of RNNs
You can specify the initial state of RNN layers symbolically by
calling them with the option initialState
. The value of
initialState
should be a tensor or list of tensors representing
the initial state of the RNN layer.
You can specify the initial state of RNN layers numerically by
calling resetStates
with the keyword argument states
. The value of
states
should be a numpy array or list of numpy arrays representing
the initial state of the RNN layer.
Note on passing external constants to RNNs
You can pass "external" constants to the cell using the constants
keyword argument of RNN.call
method. This requires that the cell.call
method accepts the same keyword argument constants
. Such constants
can be used to conditon the cell transformation on additional static inputs
(not changing over time), a.k.a an attention mechanism.
- config (Object)
-
cells
(tf.RNNCell[])
A
Array
of tf.RNNCell instances.
- config (Object)
-
layer
(RNN)
The instance of an
RNN
layer to be wrapped. -
mergeMode
(BidirectionalMergeMode)
Mode by which outputs of the forward and backward RNNs are combinied.
If
null
orundefined
, the output will not be combined, they will be returned as anArray
. Optional
This wrapper applies a layer to every temporal slice of an input.
The input should be at least 3D, and the dimension of the index 1
will be
considered to be the temporal dimension.
Consider a batch of 32 samples, where each sample is a sequence of 10 vectors
of 16 dimensions. The batch input shape of the layer is then [32, 10, 16]
, and the inputShape
, not including the sample dimension, is
[10, 16]
.
You can then use TimeDistributed
to apply a Dense
layer to each of the 10
timesteps, independently:
const model = tf.sequential();
model.add(tf.layers.timeDistributed({
layer: tf.layers.dense({units: 8}),
inputShape: [10, 16],
}));
// Now model.outputShape = [null, 10, 8].
// The output will then have shape `[32, 10, 8]`.
// In subsequent layers, there is no need for `inputShape`:
model.add(tf.layers.timeDistributed({layer: tf.layers.dense({units: 32})}));
// Now model.outputShape = [null, 10, 32].
The output will then have shape [32, 10, 32]
.
TimeDistributed
can be used with arbitrary layers, not just Dense
, for
instance a Conv2D
layer.
const model = tf.sequential();
model.add(tf.layers.timeDistributed({
layer: tf.layers.conv2d({filters: 64, kernelSize: [3, 3]}),
inputShape: [10, 299, 299, 3],
}));
- config (Object)
- layer (tf.layers.Layer) The layer to be wrapped.
A layer is a grouping of operations and weights that can be composed to create a tf.Model.
Layers are constructed by using the functions under the tf.layers namespace.
Builds or executes a `Layer's logic.
When called with tf.Tensor(s), execute the Layer
s computation and
return Tensor(s). For example:
const denseLayer = tf.layers.dense({
units: 1,
kernelInitializer: 'zeros',
useBias: false
});
// Invoke the layer's apply() method with a [tf.Tensor](#class:Tensor) (with concrete
// numeric values).
const input = tf.ones([2, 2]);
const output = denseLayer.apply(input);
// The output's value is expected to be [[0], [0]], due to the fact that
// the dense layer has a kernel initialized to all-zeros and does not have
// a bias.
output.print();
When called with tf.SymbolicTensor(s), this will prepare the layer for future execution. This entails internal book-keeping on shapes of expected Tensors, wiring layers together, and initializing weights.
Calling apply
with tf.SymbolicTensors are typically used during the
building of non-tf.Sequential models. For example:
const flattenLayer = tf.layers.flatten();
const denseLayer = tf.layers.dense({units: 1});
// Use tf.layers.input() to obtain a SymbolicTensor as input to apply().
const input = tf.input({shape: [2, 2]});
const output1 = flattenLayer.apply(input);
// output1.shape is [null, 4]. The first dimension is the undetermined
// batch size. The second dimension comes from flattening the [2, 2]
// shape.
console.log(output1.shape);
// The output SymbolicTensor of the flatten layer can be used to call
// the apply() of the dense layer:
const output2 = denseLayer.apply(output1);
// output2.shape is [null, 1]. The first dimension is the undetermined
// batch size. The second dimension matches the number of units of the
// dense layer.
console.log(output2.shape);
// The input and output and be used to construct a model that consists
// of the flatten and dense layers.
const model = tf.model({inputs: input, outputs: output2});
- inputs (tf.Tensor|tf.Tensor[]|tf.SymbolicTensor|tf.SymbolicTensor[]) a tf.Tensor or tf.SymbolicTensor or an Array of them.
-
kwargs
(any)
Additional keyword arguments to be passed to
call()
. Optional
An input layer is an entry point into a tf.Model.
InputLayer
is generated automatically for tf.Sequential models by specifying
the inputshape
or batchInputShape
for the first layer. It should not be
specified explicitly.
// Define a model which simply adds two inputs.
const inputA = tf.input({shape: [3]});
const inputB = tf.input({shape: [3]});
const sum = tf.layers.add().apply([inputA, inputB]);
const model = tf.model({inputs: [inputA, inputB], outputs: sum});
const batchSize = 2;
model.predict([tf.ones([batchSize, 3]), tf.ones([batchSize, 3])]).print();
- config (Object)
- inputShape (number[]) Input shape, not including the batch axis. Optional
- batchSize (number) Optional input batch size (integer or null). Optional
- batchInputShape (number[]) Batch input shape, including the batch axis. Optional
- dtype ('float32'|'int32'|'bool') Datatype of the input. Optional
- sparse (boolean) Whether the placeholder created is meant to be sparse. Optional
- name (string) Name of the layer. Optional
To perform mathematical computation on Tensors, we use operations. Tensors are immutable, so all operations always return new Tensors and never modify input Tensors.
Adds two tf.Tensors element-wise, A + B. Supports broadcasting.
We also expose addStrict
which has the same signature as this op and
asserts that a
and b
are the same shape (does not broadcast).
const a = tf.tensor1d([1, 2, 3, 4]);
const b = tf.tensor1d([10, 20, 30, 40]);
a.add(b).print(); // or tf.add(a, b)
// Broadcast add a with b.
const a = tf.scalar(5);
const b = tf.tensor1d([10, 20, 30, 40]);
a.add(b).print(); // or tf.add(a, b)
Subtracts two tf.Tensors element-wise, A - B. Supports broadcasting.
We also expose subStrict
which has the same signature as this op and
asserts that a
and b
are the same shape (does not broadcast).
const a = tf.tensor1d([10, 20, 30, 40]);
const b = tf.tensor1d([1, 2, 3, 4]);
a.sub(b).print(); // or tf.sub(a, b)
// Broadcast subtract a with b.
const a = tf.tensor1d([10, 20, 30, 40]);
const b = tf.scalar(5);
a.sub(b).print(); // or tf.sub(a, b)
Multiplies two tf.Tensors element-wise, A * B. Supports broadcasting.
We also expose mulStrict
which has the same signature as this op and
asserts that a
and b
are the same shape (does not broadcast).
const a = tf.tensor1d([1, 2, 3, 4]);
const b = tf.tensor1d([2, 3, 4, 5]);
a.mul(b).print(); // or tf.mul(a, b)
// Broadcast mul a with b.
const a = tf.tensor1d([1, 2, 3, 4]);
const b = tf.scalar(5);
a.mul(b).print(); // or tf.mul(a, b)
Divides two tf.Tensors element-wise, A / B. Supports broadcasting.
We also expose divStrict
which has the same signature as this op and
asserts that a
and b
are the same shape (does not broadcast).
const a = tf.tensor1d([1, 4, 9, 16]);
const b = tf.tensor1d([1, 2, 3, 4]);
a.div(b).print(); // or tf.div(a, b)
// Broadcast div a with b.
const a = tf.tensor1d([2, 4, 6, 8]);
const b = tf.scalar(2);
a.div(b).print(); // or tf.div(a, b)
Returns the max of a and b (a > b ? a : b
) element-wise.
Supports broadcasting.
We also expose maximumStrict
which has the same signature as this op and
asserts that a
and b
are the same shape (does not broadcast).
const a = tf.tensor1d([1, 4, 3, 16]);
const b = tf.tensor1d([1, 2, 9, 4]);
a.maximum(b).print(); // or tf.maximum(a, b)
// Broadcast maximum a with b.
const a = tf.tensor1d([2, 4, 6, 8]);
const b = tf.scalar(5);
a.maximum(b).print(); // or tf.maximum(a, b)
Returns the min of a and b (a < b ? a : b
) element-wise.
Supports broadcasting.
We also expose minimumStrict
which has the same signature as this op and
asserts that a
and b
are the same shape (does not broadcast).
const a = tf.tensor1d([1, 4, 3, 16]);
const b = tf.tensor1d([1, 2, 9, 4]);
a.minimum(b).print(); // or tf.minimum(a, b)
// Broadcast minimum a with b.
const a = tf.tensor1d([2, 4, 6, 8]);
const b = tf.scalar(5);
a.minimum(b).print(); // or tf.minimum(a, b)
Computes the power of one tf.Tensor to another. Supports broadcasting.
Given a tf.Tensor x and a tf.Tensor y, this operation computes x^y for corresponding elements in x and y.
const a = tf.tensor([[2, 3], [4, 5]])
const b = tf.tensor([[1, 2], [3, 0]]).toInt();
a.pow(b).print(); // or tf.pow(a, b)
const a = tf.tensor([[1, 2], [3, 4]])
const b = tf.tensor(2).toInt();
a.pow(b).print(); // or tf.pow(a, b)
We also expose powStrict
which has the same signature as this op and
asserts that base
and tf.exp() are the same shape (does not broadcast).
Computes absolute value element-wise: abs(x)
const x = tf.tensor1d([-1, 2, -3, 4]);
x.abs().print(); // or tf.abs(x)
Computes acos of the input tf.Tensor element-wise: acos(x)
const x = tf.tensor1d([0, 1, -1, .7]);
x.acos().print(); // or tf.acos(x)
- x (tf.Tensor) The input tensor.
Computes asin of the input tf.Tensor element-wise: asin(x)
const x = tf.tensor1d([0, 1, -1, .7]);
x.asin().print(); // or tf.asin(x)
- x (tf.Tensor) The input tensor.
Computes atan of the input tf.Tensor element-wise: atan(x)
const x = tf.tensor1d([0, 1, -1, .7]);
x.atan().print(); // or tf.atan(x)
- x (tf.Tensor) The input tensor.
Computes ceiling of input tf.Tensor element-wise: ceil(x)
const x = tf.tensor1d([.6, 1.1, -3.3]);
x.ceil().print(); // or tf.ceil(x)
- x (tf.Tensor) The input Tensor.
Clips values element-wise. max(min(x, clipValueMax), clipValueMin)
const x = tf.tensor1d([-1, 2, -3, 4]);
x.clipByValue(-2, 3).print(); // or tf.clipByValue(x, -2, 3)
- x (tf.Tensor) The input tensor.
- clipValueMin (number) Lower-bound of range to be clipped to.
- clipValueMax (number) Upper-bound of range to be clipped to.
Computes cos of the input tf.Tensor element-wise: cos(x)
const x = tf.tensor1d([0, Math.PI / 2, Math.PI * 3 / 4]);
x.cos().print(); // or tf.cos(x)
- x (tf.Tensor) The input tensor.
Computes hyperbolic cos of the input tf.Tensor element-wise: cosh(x)
const x = tf.tensor1d([0, 1, -1, .7]);
x.cosh().print(); // or tf.cosh(x)
- x (tf.Tensor) The input tensor.
Computes exponential linear element-wise, x > 0 ? e ^ x - 1 : 0
const x = tf.tensor1d([-1, 1, -3, 2]);
x.elu().print(); // or tf.elu(x)
- x (tf.Tensor) The input tensor.
Computes exponential of the input tf.Tensor element-wise. e ^ x
const x = tf.tensor1d([1, 2, -3]);
x.exp().print(); // or tf.exp(x)
- x (tf.Tensor) The input tensor.
Computes floor of input tf.Tensor element-wise: floor(x)
.
const x = tf.tensor1d([.6, 1.1, -3.3]);
x.floor().print(); // or tf.floor(x)
- x (tf.Tensor) The input tensor.
Computes leaky rectified linear element-wise.
See http://web.stanford.edu/~awni/papers/relu_hybrid_icml2013_final.pdf
const x = tf.tensor1d([-1, 2, -3, 4]);
x.leakyRelu(0.1).print(); // or tf.leakyRelu(x, 0.1)
- x (tf.Tensor) The input tensor.
- alpha (number) The scaling factor for negative values, defaults to 0.2. Optional
Computes natural logarithm of the input tf.Tensor element-wise: ln(x)
const x = tf.tensor1d([1, 2, Math.E]);
x.log().print(); // or tf.log(x)
- x (tf.Tensor) The input tensor.
Computes natural logarithm of the input tf.Tensor plus one
element-wise: ln(1 + x)
const x = tf.tensor1d([1, 2, Math.E - 1]);
x.log1p().print(); // or tf.log1p(x)
- x (tf.Tensor) The input tensor.
Computes -1 * x
element-wise.
const x = tf.tensor2d([1, 2, -2, 0], [2, 2]);
x.neg().print(); // or tf.neg(x)
- x (tf.Tensor) The input tensor.
Computes leaky rectified linear element-wise with parametric alphas.
x < 0 ? alpha * x : f(x) = x
const x = tf.tensor1d([-1, 2, -3, 4]);
const alpha = tf.scalar(0.1);
x.prelu(alpha).print(); // or tf.prelu(x, alpha)
Computes rectified linear element-wise: max(x, 0)
const x = tf.tensor1d([-1, 2, -3, 4]);
x.relu().print(); // or tf.relu(x)
- x (tf.Tensor) The input tensor.
Computes scaled exponential linear element-wise.
x < 0 ? scale * alpha * (exp(x) - 1) : x
const x = tf.tensor1d([-1, 2, -3, 4]);
x.selu().print(); // or tf.selu(x)
- x (tf.Tensor) The input tensor.
Computes sigmoid element-wise, 1 / (1 + exp(-x))
const x = tf.tensor1d([0, -1, 2, -3]);
x.sigmoid().print(); // or tf.sigmoid(x)
- x (tf.Tensor) The input tensor.
Computes sin of the input Tensor element-wise: sin(x)
const x = tf.tensor1d([0, Math.PI / 2, Math.PI * 3 / 4]);
x.sin().print(); // or tf.sin(x)
- x (tf.Tensor) The input tensor.
Computes hyperbolic sin of the input tf.Tensor element-wise: sinh(x)
const x = tf.tensor1d([0, 1, -1, .7]);
x.sinh().print(); // or tf.sinh(x)
- x (tf.Tensor) The input tensor.
Computes square root of the input tf.Tensor element-wise: y = sqrt(x)
const x = tf.tensor1d([1, 2, 4, -1]);
x.sqrt().print(); // or tf.sqrt(x)
- x (tf.Tensor) The input tensor.
Computes square of x
element-wise: x ^ 2
const x = tf.tensor1d([1, 2, Math.sqrt(2), -1]);
x.square().print(); // or tf.square(x)
- x (tf.Tensor) The input Tensor.
Computes step of the input tf.Tensor element-wise: x > 0 ? 1 : alpha * x
const x = tf.tensor1d([0, 2, -1, -3]);
x.step(.5).print(); // or tf.step(x, .5)
- x (tf.Tensor) The input tensor.
- alpha (number) The gradient when input is negative. Optional
Computes tan of the input tf.Tensor element-wise, tan(x)
const x = tf.tensor1d([0, Math.PI / 2, Math.PI * 3 / 4]);
x.tan().print(); // or tf.tan(x)
- x (tf.Tensor) The input tensor.
Computes hyperbolic tangent of the input tf.Tensor element-wise: tanh(x)
const x = tf.tensor1d([0, 1, -1, 70]);
x.tanh().print(); // or tf.tanh(x)
- x (tf.Tensor) The input tensor.
Computes the dot product of two matrices, A * B. These must be matrices.
const a = tf.tensor2d([1, 2], [1, 2]);
const b = tf.tensor2d([1, 2, 3, 4], [2, 2]);
a.matMul(b).print(); // or tf.matMul(a, b)
- a (tf.Tensor2D) First matrix in dot product operation.
- b (tf.Tensor2D) Second matrix in dot product operation.
-
transposeA
(boolean)
If true,
a
is transposed before multiplication. Optional -
transposeB
(boolean)
If true,
b
is transposed before multiplication. Optional
Computes the norm of scalar, vectors, and matrices. This function can compute several different vector norms (the 1-norm, the Euclidean or 2-norm, the inf-norm, and in general the p-norm for p > 0) and matrix norms (Frobenius, 1-norm, and inf-norm).
const x = tf.tensor1d([1, 2, 3, 4]);
x.norm().print(); // or tf.norm(x)
- x (tf.Tensor) The input array.
-
ord
(number|'euclidean'|'fro')
Optional. Order of the norm. Supported norm types are
following:
ord norm for matrices norm for vectors 'euclidean' Frobenius norm 2-norm 'fro' Frobenius norm Infinity max(sum(abs(x), axis=1)) max(abs(x)) -Infinity min(sum(abs(x), axis=1)) min(abs(x)) 1 max(sum(abs(x), axis=0)) sum(abs(x)) 2 sum(abs(x)^2)^1/2* - axis (number|number[]) Optional. If axis is null (the default), the input is considered a vector and a single vector norm is computed over the entire set of values in the Tensor, i.e. norm(x, ord) is equivalent to norm(x.reshape([-1]), ord). If axis is a integer, the input is considered a batch of vectors, and axis determines the axis in x over which to compute vector norms. If axis is a 2-tuple of integer it is considered a batch of matrices and axis determines the axes in NDArray over which to compute a matrix norm. Optional
- keepDims (boolean) Optional. If true, the norm have the same dimensionality as the input. Optional
Computes the outer product of two vectors, v1 and v2.
const a = tf.tensor1d([1, 2, 3]);
const b = tf.tensor1d([3, 4, 5]);
tf.outerProduct(a, b).print();
- v1 (tf.Tensor1D) The first vector in the outer product operation.
- v2 (tf.Tensor1D) The second vector in the dot product operation.
Transposes the tf.Tensor. Permutes the dimensions according to perm
.
The returned tf.Tensor's dimension i
will correspond to the input
dimension perm[i]
. If perm
is not given, it is set to [n-1...0]
,
where n
is the rank of the input tf.Tensor. Hence by default, this
operation performs a regular matrix transpose on 2-D input tf.Tensors.
const a = tf.tensor2d([1, 2, 3, 4, 5, 6], [2, 3]);
a.transpose().print(); // or tf.transpose(a)
- x (tf.Tensor) The tensor to transpose.
- perm (number[]) The permutation of the dimensions of a. Optional
Computes the 2D average pooling of an image.
-
x
(tf.Tensor3D|tf.Tensor4D)
The input tensor, of rank 4 or rank 3 of shape
[batch, height, width, inChannels]
. If rank 3, batch of 1 is assumed. -
filterSize
([number, number]|number)
The filter size, a tuple
[filterHeight, filterWidth]
. -
strides
([number, number]|number)
The strides of the pooling:
[strideHeight, strideWidth]
. -
pad
('valid'|'same'|number)
The type of padding algorithm:
same
and stride 1: output will be of same size as input, regardless of filter size.valid
: output will be smaller than input if filter is larger than 1x1.- For more info, see this guide: https://www.tensorflow.org/api_guides/python/nn#Convolution
- dimRoundingMode ('floor'|'round'|'ceil') The rounding mode used when computing output dimensions if pad is a number. If none is provided, it will not round and error if the output is of fractional size. Optional
Computes a 1D convolution over the input x.
-
input
(tf.Tensor2D|tf.Tensor3D)
The input tensor, of rank 3 or rank 2, of shape
[batch, width, inChannels]
. If rank 2, batch of 1 is assumed. -
filter
(tf.Tensor3D)
The filter, rank 3, of shape
[filterWidth, inDepth, outDepth]
. - stride (number) The number of entries by which the filter is moved right at each step.
-
pad
('valid'|'same'|number)
The type of padding algorithm.
same
and stride 1: output will be of same size as input, regardless of filter size.valid
: output will be smaller than input if filter is larger than 1x1.- For more info, see this guide: https://www.tensorflow.org/api_guides/python/nn#Convolution
- dataFormat ('NWC'|'NCW') An optional string from "NWC", "NCW". Defaults to "NWC", the data is stored in the order of [batch, in_width, in_channels]. Only "NWC" is currently supported. Optional
-
dilation
(number)
The dilation rate in which we sample input values in
atrous convolution. Defaults to
1
. If it is greater than 1, then stride must be1
. Optional - dimRoundingMode ('floor'|'round'|'ceil') The rounding mode used when computing output dimensions if pad is a number. If none is provided, it will not round and error if the output is of fractional size. Optional
Computes a 2D convolution over the input x.
-
x
(tf.Tensor3D|tf.Tensor4D)
The input tensor, of rank 4 or rank 3, of shape
[batch, height, width, inChannels]
. If rank 3, batch of 1 is assumed. -
filter
(tf.Tensor4D)
The filter, rank 4, of shape
[filterHeight, filterWidth, inDepth, outDepth]
. -
strides
([number, number]|number)
The strides of the convolution:
[strideHeight, strideWidth]
. -
pad
('valid'|'same'|number)
The type of padding algorithm.
same
and stride 1: output will be of same size as input, regardless of filter size.valid
: output will be smaller than input if filter is larger than 1x1.- For more info, see this guide: https://www.tensorflow.org/api_guides/python/nn#Convolution
- dataFormat ('NHWC'|'NCHW') : An optional string from: "NHWC", "NCHW". Defaults to "NHWC". Specify the data format of the input and output data. With the default format "NHWC", the data is stored in the order of: [batch, height, width, channels]. Only "NHWC" is currently supported. Optional
-
dilations
([number, number]|number)
The dilation rates:
[dilationHeight, dilationWidth]
in which we sample input values across the height and width dimensions in atrous convolution. Defaults to[1, 1]
. Ifdilations
is a single number, thendilationHeight == dilationWidth
. If it is greater than 1, then all values ofstrides
must be 1. Optional - dimRoundingMode ('floor'|'round'|'ceil') The rounding mode used when computing output dimensions if pad is a number. If none is provided, it will not round and error if the output is of fractional size. Optional
Computes the transposed 2D convolution of an image, also known as a deconvolution.
-
x
(tf.Tensor3D|tf.Tensor4D)
The input image, of rank 4 or rank 3, of shape
[batch, height, width, inDepth]
. If rank 3, batch of 1 is assumed. -
filter
(tf.Tensor4D)
The filter, rank 4, of shape
[filterHeight, filterWidth, outDepth, inDepth]
.inDepth
must matchinDepth
inx
. -
outputShape
([number, number, number, number]|[number, number, number])
Output shape, of rank 4 or rank 3:
[batch, height, width, outDepth]
. If rank 3, batch of 1 is assumed. -
strides
([number, number]|number)
The strides of the original convolution:
[strideHeight, strideWidth]
. - pad ('valid'|'same'|number) The type of padding algorithm used in the non-transpose version of the op.
- dimRoundingMode ('floor'|'round'|'ceil') The rounding mode used when computing output dimensions if pad is a number. If none is provided, it will not round and error if the output is of fractional size. Optional
Depthwise 2D convolution.
Given a 4D tf.input() array and a filter
array of shape
[filterHeight, filterWidth, inChannels, channelMultiplier]
containing
inChannels
convolutional filters of depth 1, this op applies a
different filter to each input channel (expanding from 1 channel to
channelMultiplier
channels for each), then concatenates the results
together. The output has inChannels * channelMultiplier
channels.
See https://www.tensorflow.org/api_docs/python/tf/nn/depthwise_conv2d for more details.
-
input
(tf.Tensor3D|tf.Tensor4D)
The input tensor, of rank 4 or rank 3, of shape
[batch, height, width, inChannels]
. If rank 3, batch of 1 is assumed. -
filter
(tf.Tensor4D)
The filter tensor, rank 4, of shape
[filterHeight, filterWidth, inChannels, channelMultiplier]
. -
strides
([number, number]|number)
The strides of the convolution:
[strideHeight, strideWidth]
. If strides is a single number, thenstrideHeight == strideWidth
. -
pad
('valid'|'same'|number)
The type of padding algorithm.
same
and stride 1: output will be of same size as input, regardless of filter size.valid
: output will be smaller than input if filter is larger than 1x1.- For more info, see this guide: https://www.tensorflow.org/api_guides/python/nn#Convolution
- dataFormat ('NHWC'|'NCHW') : An optional string from: "NHWC", "NCHW". Defaults to "NHWC". Specify the data format of the input and output data. With the default format "NHWC", the data is stored in the order of: [batch, height, width, channels]. Only "NHWC" is currently supported. Optional
-
dilations
([number, number]|number)
The dilation rates:
[dilationHeight, dilationWidth]
in which we sample input values across the height and width dimensions in atrous convolution. Defaults to[1, 1]
. Ifrate
is a single number, thendilationHeight == dilationWidth
. If it is greater than 1, then all values ofstrides
must be 1. Optional - dimRoundingMode ('floor'|'round'|'ceil') The rounding mode used when computing output dimensions if pad is a number. If none is provided, it will not round and error if the output is of fractional size. Optional
Computes the 2D max pooling of an image.
-
x
(tf.Tensor3D|tf.Tensor4D)
The input tensor, of rank 4 or rank 3 of shape
[batch, height, width, inChannels]
. If rank 3, batch of 1 is assumed. -
filterSize
([number, number]|number)
The filter size, a tuple
[filterHeight, filterWidth]
. -
strides
([number, number]|number)
The strides of the pooling:
[strideHeight, strideWidth]
. -
pad
('valid'|'same'|number)
The type of padding algorithm.
same
and stride 1: output will be of same size as input, regardless of filter size.valid
: output will be smaller than input if filter is larger than 1x1.- For more info, see this guide: https://www.tensorflow.org/api_guides/python/nn#Convolution
- dimRoundingMode ('floor'|'round'|'ceil') The rounding mode used when computing output dimensions if pad is a number. If none is provided, it will not round and error if the output is of fractional size. Optional
Computes the 2D min pooling of an image.
-
input
(tf.Tensor3D|tf.Tensor4D)
The input tensor, of rank 4 or rank 3 of shape
[batch, height, width, inChannels]
. If rank 3, batch of 1 is assumed. -
filterSize
([number, number]|number)
The filter size, a tuple
[filterHeight, filterWidth]
. -
strides
([number, number]|number)
The strides of the pooling:
[strideHeight, strideWidth]
. -
pad
('valid'|'same'|number)
The type of padding algorithm.
same
and stride 1: output will be of same size as input, regardless of filter size.valid
: output will be smaller than input if filter is larger than 1x1.- For more info, see this guide: https://www.tensorflow.org/api_guides/python/nn#Convolution
- dimRoundingMode ('floor'|'round'|'ceil') The rounding mode used when computing output dimensions if pad is a number. If none is provided, it will not round and error if the output is of fractional size. Optional
Returns the indices of the maximum values along an axis
.
The result has the same shape as tf.input() with the dimension along axis
removed.
const x = tf.tensor1d([1, 2, 3]);
x.argMax().print(); // or tf.argMax(x)
const x = tf.tensor2d([1, 2, 4, 3], [2, 2]);
const axis = 1;
x.argMax(axis).print(); // or tf.argMax(x, axis)
- x (tf.Tensor) The input tensor.
- axis (number) The dimension to reduce. By default it reduces across all axes and returns the flat index Optional
Returns the indices of the minimum values along an axis
.
The result has the same shape as tf.input() with the dimension along axis
removed.
const x = tf.tensor1d([1, 2, 3]);
x.argMin().print(); // or tf.argMin(x)
const x = tf.tensor2d([1, 2, 4, 3], [2, 2]);
const axis = 1;
x.argMin(axis).print(); // or tf.argMin(x, axis)
- x (tf.Tensor) The input tensor.
- axis (number) The dimension to reduce. By default it reduces across all axes and returns the flat index. Optional
Computes the log(sum(exp(elements across the reduction dimensions)).
Reduces the input along the dimensions given in axis
. Unless keepDims
is true, the rank of the array is reduced by 1 for each entry in axis
.
If keepDims
is true, the reduced dimensions are retained with length 1.
If axis
has no entries, all dimensions are reduced, and an array with a
single element is returned.
const x = tf.tensor1d([1, 2, 3]);
x.logSumExp().print(); // or tf.logSumExp(x)
const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const axis = 1;
x.logSumExp(axis).print(); // or tf.logSumExp(a, axis)
- input (tf.Tensor) The input tensor.
- axis (number|number[]) The dimension(s) to reduce. If null (the default), reduces all dimensions. Optional
- keepDims (boolean) If true, retains reduced dimensions with length of 1. Defaults to false. Optional
Computes the maximum of elements across dimensions of a tf.Tensor.
Reduces the input along the dimensions given in axes
. Unless keepDims
is true, the rank of the tf.Tensor is reduced by 1 for each entry in axes
.
If keepDims
is true, the reduced dimensions are retained with length 1.
If axes
has no entries, all dimensions are reduced, and an tf.Tensor with
a single element is returned.
const x = tf.tensor1d([1, 2, 3]);
x.max().print(); // or tf.max(x)
const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const axis = 1;
x.max(axis).print(); // or tf.max(x, axis)
- x (tf.Tensor) The input tensor.
- axis (number|number[]) The dimension(s) to reduce. By default it reduces all dimensions. Optional
- keepDims (boolean) If true, retains reduced dimensions with size 1. Optional
Computes the mean of elements across dimensions of a tf.Tensor.
Reduces x
along the dimensions given in axis
. Unless keepDims
is
true, the rank of the tf.Tensor is reduced by 1 for each entry in axis
.
If keepDims
is true, the reduced dimensions are retained with length 1.
If axis
has no entries, all dimensions are reduced, and a tf.Tensor with
a single element is returned.
const x = tf.tensor1d([1, 2, 3]);
x.mean().print(); // or tf.logSumExp(a)
const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const axis = 1;
x.mean(axis).print(); // or tf.mean(x, axis)
- x (tf.Tensor) The input tensor.
- axis (number|number[]) The dimension(s) to reduce. By default it reduces all dimensions. Optional
- keepDims (boolean) If true, retains reduced dimensions with size 1. Optional
Computes the minimum value from the input.
Reduces the input along the dimensions given in axes
. Unless keepDims
is true, the rank of the array is reduced by 1 for each entry in axes
.
If keepDims
is true, the reduced dimensions are retained with length 1.
If axes
has no entries, all dimensions are reduced, and an array with a
single element is returned.
const x = tf.tensor1d([1, 2, 3]);
x.min().print(); // or tf.min(x)
const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const axis = 1;
x.min(axis).print(); // or tf.min(x, axis)
- x (tf.Tensor) The input Tensor.
- axis (number|number[]) The dimension(s) to reduce. By default it reduces all dimensions. Optional
- keepDims (boolean) If true, retains reduced dimensions with size 1. Optional
Computes the sum of elements across dimensions of a tf.Tensor.
Reduces the input along the dimensions given in axes
. Unless keepDims
is true, the rank of the tf.Tensor is reduced by 1 for each entry in axes
.
If keepDims
is true, the reduced dimensions are retained with length 1.
If axes has no entries, all dimensions are reduced, and a tf.Tensor with a
single element is returned.
const x = tf.tensor1d([1, 2, 3]);
x.sum().print(); // or tf.logSumExp(x)
const x = tf.tensor2d([1, 2, 3, 4], [2, 2]);
const axis = 1;
x.sum(axis).print(); // or tf.sum(x, axis)
- x (tf.Tensor) The input tensor to compute the sum over.
- axis (number|number[]) The dimension(s) to reduce. By default it reduces all dimensions. Optional
- keepDims (boolean) If true, retains reduced dimensions with size 1. Optional
Batch normalization.
As described in http://arxiv.org/abs/1502.03167.
Mean, variance, scale, and offset can be of two shapes:
- The same shape as the input.
- In the common case, the depth dimension is the last dimension of x, so the values would be an tf.Tensor1D of shape [depth].
- x (tf.Tensor) The input Tensor.
- mean (tf.Tensor|tf.Tensor1D) A mean Tensor.
- variance (tf.Tensor|tf.Tensor1D) A variance Tensor.
- varianceEpsilon (number) A small float number to avoid dividing by 0. Optional
- scale (tf.Tensor|tf.Tensor1D) A scale Tensor. Optional
- offset (tf.Tensor|tf.Tensor1D) An offset Tensor. Optional
Normalizes the activation of a local neighborhood across or within channels.
- x (tf.Tensor3D|tf.Tensor4D) The input tensor. The 4-D input tensor is treated as a 3-D array of 1D vectors (along the last dimension), and each vector is normalized independently.
- radius (number) The number of adjacent channels or spatial locations of the 1D normalization window. In Tensorflow this param is called 'depth_radius' because only 'acrossChannels' mode is supported. Optional
- bias (number) A constant bias term for the basis. Optional
- alpha (number) A scale factor, usually positive. Optional
- beta (number) An exponent. Optional
- normRegion ('acrossChannels'|'withinChannel') Default is 'acrossChannels'. Optional
Calculates the mean and variance of x
. The mean and variance are
calculated by aggregating the contents of x
across axes
. If x
is
1-D and axes = [0]
this is just the mean and variance of a vector.
- x (tf.Tensor) The input tensor.
- axis (number|number[]) The dimension(s) along with to compute mean and variance. By default it reduces all dimensions. Optional
- keepDims (boolean) If true, the moments have the same dimensionality as the input. Optional
Computes the softmax normalized vector given the logits.
const a = tf.tensor1d([1, 2, 3]);
a.softmax().print(); // or tf.softmax(a)
const a = tf.tensor2d([2, 4, 6, 1, 2, 3], [2, 3]);
a.softmax().print(); // or tf.softmax(a)
- logits (tf.Tensor) The logits array.
-
dim
(number)
The dimension softmax would be performed on. Defaults to
-1
which indicates the last dimension. Optional
Bilinear resize a batch of 3D images to a new shape.
-
images
(tf.Tensor3D|tf.Tensor4D)
The images, of rank 4 or rank 3, of shape
[batch, height, width, inChannels]
. If rank 3, batch of 1 is assumed. -
size
([number, number])
The new shape
[newHeight, newWidth]
to resize the images to. Each channel is resized individually. -
alignCorners
(boolean)
Defaults to False. If true, rescale
input by
(new_height - 1) / (height - 1)
, which exactly aligns the 4 corners of images and resized images. If false, rescale bynew_height / height
. Treat similarly the width dimension. Optional
Computes the next state and output of a BasicLSTMCell.
Returns [newC, newH]
.
Derived from tf.contrib.rnn.BasicLSTMCell.
- forgetBias (tf.Scalar) Forget bias for the cell.
- lstmKernel (tf.Tensor2D) The weights for the cell.
- lstmBias (tf.Tensor1D) The bias for the cell.
- data (tf.Tensor2D) The input to the cell.
- c (tf.Tensor2D) Previous cell state.
- h (tf.Tensor2D) Previous cell output.
Computes the next states and outputs of a stack of LSTMCells.
Each cell output is used as input to the next cell.
Returns [cellState, cellOutput]
.
Derived from tf.contrib.rn.MultiRNNCell.
- lstmCells ((data: tf.Tensor2D, c: tf.Tensor2D, h: tf.Tensor2D): [tf.Tensor2D, tf.Tensor2D][]) Array of LSTMCell functions.
- data (tf.Tensor2D) The input to the cell.
- c (tf.Tensor2D[]) Array of previous cell states.
- h (tf.Tensor2D[]) Array of previous cell outputs.
Returns the truth value of (a == b) element-wise. Supports broadcasting.
We also expose equalStrict
which has the same signature as this op
and asserts that a
and b
are the same shape (does not broadcast).
Returns the truth value of (a > b) element-wise. Supports broadcasting.
We also expose greaterStrict
which has the same signature as this
op and asserts that a
and b
are the same shape (does not broadcast).
Returns the truth value of (a >= b) element-wise. Supports broadcasting.
We also expose greaterStrict
which has the same signature as this
op and asserts that a
and b
are the same shape (does not broadcast).
Returns the truth value of (a < b) element-wise. Supports broadcasting.
We also expose lessStrict
which has the same signature as this op and
asserts that a
and b
are the same shape (does not broadcast).
Returns the truth value of (a <= b) element-wise. Supports broadcasting.
We also expose lessEqualStrict
which has the same signature as this op
and asserts that a
and b
are the same shape (does not broadcast).
Returns the truth value of a AND b element-wise. Supports broadcasting.
Returns the truth value of NOT x
element-wise.
- x (tf.Tensor) The input tensor. Must be of dtype 'bool'.
Returns the truth value of a OR b
element-wise. Supports broadcasting.
Returns the truth value of a XOR b
element-wise. Supports broadcasting.
Returns the truth value of (a != b) element-wise. Supports broadcasting.
We also expose notEqualStrict
which has the same signature as this op and
asserts that a
and b
are the same shape (does not broadcast).
Returns the elements, either a
or b
depending on the condition
.
If the condition is true, select from a
, otherwise select from b
.
We also provide an API to do perform training, and compute gradients. We compute gradients eagerly, users provide a function that is a combination of operations and we automatically differentiate that function's output with respect to its inputs.
For those familiar with TensorFlow, the API we expose exactly mirrors the TensorFlow Eager API.
Provided f(x)
, returns another function g(x, dy?)
, which gives the
gradient of f(x)
with respect to x
.
If dy
is provided, the gradient of f(x).mul(dy).sum()
with respect to
x
is computed instead. f(x)
must take a single tensor x
and return a
single tensor y
. If f()
takes multiple inputs, use tf.grads() instead.
// f(x) = x ^ 2
const f = x => x.square();
// f'(x) = 2x
const g = tf.grad(f);
const x = tf.tensor1d([2, 3]);
g(x).print();
// f(x) = x ^ 3
const f = x => x.pow(tf.scalar(3, 'int32'));
// f'(x) = 3x ^ 2
const g = tf.grad(f);
// f''(x) = 6x
const gg = tf.grad(g);
const x = tf.tensor1d([2, 3]);
gg(x).print();
Provided f(x1, x2,...)
, returns another function g([x1, x2,...], dy?)
,
which gives an array of gradients of f()
with respect to each input
[x1
,x2
,...].
If dy
is passed when calling g()
, the gradient of
f(x1,...).mul(dy).sum()
with respect to each input is computed instead.
The provided f
must take one or more tensors and return a single tensor
y
. If f()
takes a single input, we recommend using tf.grad() instead.
// f(a, b) = a * b
const f = (a, b) => a.mul(b);
// df / da = b, df / db = a
const g = tf.grads(f);
const a = tf.tensor1d([2, 3]);
const b = tf.tensor1d([-2, -3]);
const [da, db] = g([a, b]);
console.log('da');
da.print();
console.log('db');
db.print();
Overrides the gradient computation of a function f
.
Takes a function
f(...inputs) => {value: Tensor, gradFunc: dy => Tensor[]}
and returns
another function g(...inputs)
which takes the same inputs as f
. When
called, g
returns f().value
. In backward mode, custom gradients with
respect to each input of f
are computed using f().gradFunc
.
const customOp = tf.customGrad(x => {
// Override gradient of our custom x ^ 2 op to be dy * abs(x);
return {value: x.square(), gradFunc: dy => [dy.mul(x.abs())]};
});
const x = tf.tensor1d([-1, -2, 3]);
const dx = tf.grad(x => customOp(x));
console.log(`f(x):`);
customOp(x).print();
console.log(`f'(x):`);
dx(x).print();
Like tf.grad(), but also returns the value of f()
. Useful when f()
returns a metric you want to show.
The result is a rich object with the following properties:
- grad: The gradient of
f(x)
w.r.tx
(result of tf.grad()). - value: The value returned by
f(x)
.
// f(x) = x ^ 2
const f = x => x.square();
// f'(x) = 2x
const g = tf.valueAndGrad(f);
const x = tf.tensor1d([2, 3]);
const {value, grad} = g(x);
console.log('value');
value.print();
console.log('grad');
grad.print();
Like tf.grads(), but returns also the value of f()
. Useful when f()
returns a metric you want to show.
The result is a rich object with the following properties:
- grads: The gradients of
f()
w.r.t each input (result of tf.grads()). - value: The value returned by
f(x)
.
// f(a, b) = a * b
const f = (a, b) => a.mul(b);
// df/da = b, df/db = a
const g = tf.valueAndGrads(f);
const a = tf.tensor1d([2, 3]);
const b = tf.tensor1d([-2, -3]);
const {value, grads} = g([a, b]);
const [da, db] = grads;
console.log('value');
value.print();
console.log('da');
da.print();
console.log('db');
db.print();
Computes and returns the gradient of f(x) with respect to the list of
trainable variables provided by varList
. If no list is provided, it
defaults to all trainable variables.
const a = tf.variable(tf.tensor1d([3, 4]));
const b = tf.variable(tf.tensor1d([5, 6]));
const x = tf.tensor1d([1, 2]);
// f(a, b) = a * x ^ 2 + b * x
const f = () => a.mul(x.square()).add(b.mul(x)).sum();
// df/da = x ^ 2, df/db = x
const {value, grads} = tf.variableGrads(f);
Object.keys(grads).forEach(varName => grads[varName].print());
- f (() => tf.Scalar) The function to execute. f() should return a scalar.
- varList (tf.Variable[]) The list of trainable variables. Defaults to all variables. Optional
Constructs a tf.SGDOptimizer that uses stochastic gradient descent.
// Fit a quadratic function by learning the coefficients a, b, c.
const xs = tf.tensor1d([0, 1, 2, 3]);
const ys = tf.tensor1d([1.1, 5.9, 16.8, 33.9]);
const a = tf.scalar(Math.random()).variable();
const b = tf.scalar(Math.random()).variable();
const c = tf.scalar(Math.random()).variable();
// y = a * x^2 + b * x + c.
const f = x => a.mul(x.square()).add(b.mul(x)).add(c);
const loss = (pred, label) => pred.sub(label).square().mean();
const learningRate = 0.01;
const optimizer = tf.train.sgd(learningRate);
// Train the model.
for (let i = 0; i < 10; i++) {
optimizer.minimize(() => loss(f(xs), ys));
}
// Make predictions.
console.log(
`a: ${a.dataSync()}, b: ${b.dataSync()}, c: ${c.dataSync()}`);
const preds = f(xs).dataSync();
preds.forEach((pred, i) => {
console.log(`x: ${i}, pred: ${pred}`);
});
- learningRate (number) The learning rate to use for the SGD algorithm.
Constructs a tf.MomentumOptimizer that uses momentum gradient descent.
- learningRate (number) The learning rate to use for the Momentum gradient descent algorithm.
- momentum (number) The momentum to use for the momentum gradient descent algorithm.
- useNesterov (boolean) Optional
Constructs a tf.AdagradOptimizer that uses the Adagrad algorithm. See http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf or http://ruder.io/optimizing-gradient-descent/index.html#adagrad
- learningRate (number) The learning rate to use for the Adagrad gradient descent algorithm.
- initialAccumulatorValue (number) Starting value for the accumulators, must be positive. Optional
Constructs a tf.AdadeltaOptimizer that uses the Adadelta algorithm. See https://arxiv.org/abs/1212.5701
- learningRate (number) The learning rate to use for the Adadelta gradient descent algorithm. Optional
- rho (number) The learning rate decay over each update. Optional
- epsilon (number) A constant epsilon used to better condition the grad update. Optional
Constructs a AdamOptimizer
that uses the Adam algorithm.
See https://arxiv.org/abs/1412.6980
- learningRate (number) The learning rate to use for the Adam gradient descent algorithm. Optional
- beta1 (number) The exponential decay rate for the 1st moment estimates. Optional
- beta2 (number) The exponential decay rate for the 2nd moment estimates. Optional
- epsilon (number) A small constant for numerical stability. Optional
Constructs a AdamaxOptimizer
that uses the Adamax algorithm.
See https://arxiv.org/abs/1412.6980
- learningRate (number) The learning rate to use for the Adamax gradient descent algorithm. Optional
- beta1 (number) The exponential decay rate for the 1st moment estimates. Optional
- beta2 (number) The exponential decay rate for the 2nd moment estimates. Optional
- epsilon (number) A small constant for numerical stability. Optional
- decay (number) The learning rate decay over each update. Optional
Constructs a tf.RMSPropOptimizer that uses RMSProp gradient descent. This implementation uses plain momentum and is not centered version of RMSProp.
See http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf
- learningRate (number) The learning rate to use for the RMSProp gradient descent algorithm.
- decay (number) The discounting factor for the history/coming gradient. Optional
- momentum (number) The momentum to use for the RMSProp gradient descent algorithm. Optional
- epsilon (number) Small value to avoid zero denominator. Optional
Computes softmax cross entropy between logits and labels.
Measures the probability error in discrete classification tasks in which the classes are mutually exclusive (each entry is in exactly one class). For example, each CIFAR-10 image is labeled with one and only one label: an image can be a dog or a truck, but not both.
NOTE
: While the classes are mutually exclusive, their probabilities need
not be. All that is required is that each row of labels is a valid
probability distribution. If they are not, the computation of the gradient
will be incorrect.
WARNING
: This op expects unscaled logits, since it performs a softmax on
logits internally for efficiency. Do not call this op with the output of
softmax, as it will produce incorrect results.
logits and labels must have the same shape, e.g. [batch_size, num_classes] and the same dtype.
Executes f()
and minimizes the scalar output of f()
by computing
gradients of y with respect to the list of trainable variables provided by
varList
. If no list is provided, it defaults to all trainable variables.
- f (() => tf.Scalar) The function to execute and whose output to minimize.
-
returnCost
(boolean)
Whether to return the scalar cost value produced by
executing
f()
. Optional - varList (tf.Variable[]) An optional list of variables to update. If specified, only the trainable variables in varList will be updated by minimize. Defaults to all trainable variables. Optional
Executes the provided function f
and after it is executed, cleans up all
intermediate tensors allocated by f
except those returned by f
.
f
must not return a Promise (async functions not allowed).
The returned result can be a complex object, however tidy only walks the
top-level properties (depth 1) of that object to search for tensors, or
lists of tensors that need to be tracked in the parent scope.
Using this method helps avoid memory leaks. In general, wrap calls to operations in tf.tidy() for automatic memory cleanup.
When in safe mode, you must enclose all tf.Tensor creation and ops inside a tf.tidy() to prevent memory leaks.
// y = 2 ^ 2 + 1
const y = tf.tidy(() => {
// a, b, and one will be cleaned up when the tidy ends.
const one = tf.scalar(1);
const a = tf.scalar(2);
const b = a.square();
console.log('numTensors (in tidy): ' + tf.memory().numTensors);
// The value returned inside the tidy function will return
// through the tidy, in this case to the variable y.
return b.add(one);
});
console.log('numTensors (outside tidy): ' + tf.memory().numTensors);
y.print();
- nameOrFn (string|Function) The name of the closure, or the function to execute. If a name is provided, the 2nd argument should be the function. If debug mode is on, the timing and the memory usage of the function will be tracked and displayed on the console using the provided name.
- fn (Function) The function to execute. Optional
- gradMode (boolean) If true, starts a tape and doesn't dispose tensors. Optional
Keeps a tf.Tensor generated inside a tf.tidy() from being disposed automatically.
let b;
const y = tf.tidy(() => {
const one = tf.scalar(1);
const a = tf.scalar(2);
// b will not be cleaned up by the tidy. a and one will be cleaned up
// when the tidy ends.
b = tf.keep(a.square());
console.log('numTensors (in tidy): ' + tf.memory().numTensors);
// The value returned inside the tidy function will return
// through the tidy, in this case to the variable y.
return b.add(one);
});
console.log('numTensors (outside tidy): ' + tf.memory().numTensors);
console.log('y:');
y.print();
console.log('b:');
b.print();
- result (tf.Tensor) The tensor to keep from being disposed.
Returns memory info at the current time in the program. The result is an object with the following properties:
numBytes
: Number of bytes allocated (undisposed) at this time.numTensors
: Number of unique tensors allocated.numDataBuffers
: Number of unique data buffers allocated (undisposed) at this time, which is ≤ the number of tensors (e.g.a.reshape(newShape)
makes a new Tensor that shares the same data buffer witha
).unreliable
:Optional
boolean
:- On WebGL, not present (always reliable).
- On CPU, true. Due to automatic garbage collection, these numbers
represent undisposed tensors, i.e. not wrapped in
tidy()
, or lacking a call totensor.dispose()
.
Executes f()
and returns a promise that resolves with timing
information.
The result is an object with the following properties:
wallMs
: Wall execution time.kernelMs
: Kernel execution time, ignoring data transfer.- On
WebGL
The following additional properties exist:uploadWaitMs
: CPU blocking time on texture uploads.downloadWaitMs
: CPU blocking time on texture downloads (readPixels).
const x = tf.randomNormal([20, 20]);
const time = await tf.time(() => x.matMul(x));
console.log(`kernelMs: ${time.kernelMs}, wallTimeMs: ${time.wallMs}`);
- f (() => void) The function to execute and time.
Returns a promise that resolve when a requestAnimationFrame has completed.
This is simply a sugar method so that users can do the following:
await tf.nextFrame();
TensorFlow.js can run mathematical operations on different backends. Currently, we support WebGL and JavaScript CPU. By default, we choose the 'best' backend available, but allow users to customize their backend.
Sets the backend (cpu, webgl, etc) responsible for creating tensors and executing operations on those tensors.
-
backendType
('webgl'|'cpu')
The backend type. Currently supports
'webgl'|'cpu'
. -
safeMode
(boolean)
Defaults to false. In safe mode, you are forced to
construct tensors and call math operations inside a
tidy()
which will automatically clean up intermediate tensors. Optional
Returns the current backend (cpu, webgl, etc). The backend is responsible for creating tensors and executing operations on those tensors.
Constraints are added to attributes of a Layer (such as weights, kernels, or biases) at construction time to clamp, or otherwise enforce an allowed range, of values for different components of the Layer.
MaxNorm weight constraint.
Constrains the weights incident to each hidden unit to have a norm less than or equal to a desired value.
References - Dropout: A Simple Way to Prevent Neural Networks from Overfitting Srivastava, Hinton, et al. 2014
- config (Object)
- maxValue (number) Maximum norm for incoming weights Optional
-
axis
(number)
Axis along which to calculate norms.
For instance, in a
OptionalDense
layer the weight matrix has shape[inputDim, outputDim]
, setaxis
to0
to constrain each weight vector of length[inputDim,]
. In aConv2D
layer withdataFormat="channels_last"
, the weight tensor has shape[rows, cols, inputDepth, outputDepth]
, setaxis
to[0, 1, 2]
to constrain the weights of each filter tensor of size[rows, cols, inputDepth]
.
- config (Object)
- minValue (number) Minimum norm for incoming weights Optional
- maxValue (number) Maximum norm for incoming weights Optional
-
axis
(number)
Axis along which to calculate norms.
For instance, in a
Dense
layer the weight matrix has shape[inputDim, outputDim]
, setaxis
to0
to constrain each weight vector of length[inputDim,]
. In aConv2D
layer withdataFormat="channels_last"
, the weight tensor has shape[rows, cols, inputDepth, outputDepth]
, setaxis
to[0, 1, 2]
to constrain the weights of each filter tensor of size[rows, cols, inputDepth]
. Optional -
rate
(number)
Rate for enforcing the constraint: weights will be rescaled to yield:
(1 - rate) * norm + rate * norm.clip(minValue, maxValue)
. Effectively, this means that rate=1.0 stands for strict enforcement of the constraint, while rate<1.0 means that weights will be rescaled at each step to slowly move towards a value inside the desired interval. Optional
Constains the weight to be non-negative.
Constrains the weights incident to each hidden unit to have unit norm.
- config (Object)
-
axis
(number)
Axis along which to calculate norms.
For instance, in a
OptionalDense
layer the weight matrix has shape[inputDim, outputDim]
, setaxis
to0
to constrain each weight vector of length[inputDim,]
. In aConv2D
layer withdataFormat="channels_last"
, the weight tensor has shape [rows, cols, inputDepth, outputDepth], set
axisto
[0, 1, 2]to constrain the weights of each filter tensor of size
[rows, cols, inputDepth]`.
Base class for functions that impose constraints on weight values
Initializers are used in Layers to establish the starting the values of weights, biases, kernels, etc.
Initializer that generates values initialized to some constant.
- config (Object)
- value (number) The value for each element in the variable.
Glorot normal initializer, also called Xavier normal initializer.
It draws samples from a truncated normal distribution centered on 0
with stddev = sqrt(2 / (fan_in + fan_out))
where fan_in
is the number of input units in the weight tensor
and fan_out
is the number of output units in the weight tensor.
Reference: Glorot & Bengio, AISTATS 2010 http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf
- config (Object)
- seed (number) Random number generator seed. Optional
Glorot uniform initializer, also called Xavier uniform initializer.
It draws samples from a uniform distribution within [-limit, limit]
where limit
is sqrt(6 / (fan_in + fan_out))
where fan_in
is the number of input units in the weight tensor
and fan_out
is the number of output units in the weight tensor
Reference: Glorot & Bengio, AISTATS 2010 http://jmlr.org/proceedings/papers/v9/glorot10a/glorot10a.pdf.
- config (Object)
- seed (number) Random number generator seed. Optional
He normal initializer.
It draws samples from a truncated normal distribution centered on 0
with stddev = sqrt(2 / fanIn)
where fanIn
is the number of input units in the weight tensor.
Reference: He et al., http://arxiv.org/abs/1502.01852
- config (Object)
- seed (number) Random number generator seed. Optional
Initializer that generates the identity matrix. Only use for square 2D matrices.
- config (Object)
- gain (number) Multiplicative factor to apply to the identity matrix. Optional
LeCun normal initializer.
It draws samples from a truncated normal distribution centered on 0
with stddev = sqrt(1 / fanIn)
where fanIn
is the number of input units in the weight tensor.
References: Self-Normalizing Neural Networks Efficient Backprop
- config (Object)
- seed (number) Random number generator seed. Optional
Initializer that generates tensors initialized to 1.
Initializer that generates a random orthogonal matrix.
Reference: Saxe et al., http://arxiv.org/abs/1312.6120
- config (Object)
- gain (number) Multiplicative factor to apply to the orthogonal matrix. Defaults to 1. Optional
Initializer that generates random values initialized to a normal distribution.
- config (Object)
- mean (number) Mean of the random values to generate. Optional
- stddev (number) Standard deviation of the random values to generate. Optional
- seed (number) Used to seed the random generator. Optional
Initializer that generates random values initialized to a uniform distribution.
Values will be distributed uniformly between the configured minval and maxval.
- config (Object)
- minval (number) Lower bound of the range of random values to generate. Optional
- maxval (number) Upper bound of the range of random values to generate. Optional
- seed (number) Used to seed the random generator. Optional
Initializer that generates random values initialized to a truncated normal. distribution.
These values are similar to values from a RandomNormal
except that values
more than two standard deviations from the mean are discarded and re-drawn.
This is the recommended initializer for neural network weights and filters.
- config (Object)
- mean (number) Mean of the random values to generate. Optional
- stddev (number) Standard deviation of the random values to generate. Optional
- seed (number) Used to seed the random generator. Optional
Initializer capable of adapting its scale to the shape of weights.
With distribution=NORMAL, samples are drawn from a truncated normal
distribution centered on zero, with stddev = sqrt(scale / n)
where n is:
- number of input units in the weight tensor, if mode = FAN_IN.
- number of output units, if mode = FAN_OUT.
- average of the numbers of input and output units, if mode = FAN_AVG.
With distribution=UNIFORM,
samples are drawn from a uniform distribution
within [-limit, limit], with
limit = sqrt(3 * scale / n)
.
- config (Object)
- scale (number) Scaling factor (positive float).
- mode ('fanIn'|'fanOut'|'fanAvg') Fanning mode for inputs and outputs.
- distribution ('normal'|'uniform') Probabilistic distribution of the values.
- seed (number) Random number generator seed. Optional
Initializer that generates tensors initialized to 0.
Regularizers can be attached to various components of a Layer to add a 'scoring' function to help drive weights, or other trainable values, away from excessively large values. They're typically used to promote a notion that a 'simpler' model is better than a complicated model, assuming equal performance.
Regularizer for L1 and L2 regularization.
Adds a term to the loss to penalize large weights: loss += sum(l1 * abs(x)) + sum(l2 * x^2)
Regularizer for L1 and L2 regularization.
Adds a term to the loss to penalize large weights: loss += sum(l1 * abs(x)) + sum(l2 * x^2)
- config (L1Config) Optional
Regularizer for L1 and L2 regularization.
Adds a term to the loss to penalize large weights: loss += sum(l1 * abs(x)) + sum(l2 * x^2)
- config (L1L2Config) Optional
Regularizer for L1 and L2 regularization.
Adds a term to the loss to penalize large weights: loss += sum(l1 * abs(x)) + sum(l2 * x^2)
- config (L2Config) Optional