Tensors Reference

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 an input shape defined.
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(JSON.stringify(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(JSON.stringify(model.outputs[0].shape));

You can also use an Array of already-constructed Layers 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(JSON.stringify(model.outputs[0].shape));

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().

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.

Load a model, including its topology and optionally weights. See the Tutorial named "How to import a Keras Model" for usage examples.

Example 1: Save tf.model()'s topology and weights to browser local storage; then load it back.

const model = tf.sequential(
     {layers: [tf.layers.dense({units: 1, inputShape: [3]})]});
console.log('Prediction from original model:');
model.predict(tf.ones([1, 3])).print();

const saveResults = await model.save('localstorage://my-model-1');

const loadedModel = await tf.loadModel('localstorage://my-model-1');
console.log('Prediction from loaded model:');
loadedModel.predict(tf.ones([1, 3])).print();

Example 2. Saving tf.model()'s topology and weights to browser IndexedDB; then load it back.

const model = tf.sequential(
     {layers: [tf.layers.dense({units: 1, inputShape: [3]})]});
console.log('Prediction from original model:');
model.predict(tf.ones([1, 3])).print();

const saveResults = await model.save('indexeddb://my-model-1');

const loadedModel = await tf.loadModel('indexeddb://my-model-1');
console.log('Prediction from loaded model:');
loadedModel.predict(tf.ones([1, 3])).print();

Example 3. Load a model from user-selected files from HTML file input elements.

// Note: this code snippet will not work without the HTML elements in the
//   page
const jsonUpload = document.getElementById('json-upload');
const weightsUpload = document.getElementById('weights-upload');

const model = await tf.loadModel(
     tf.io.browserFiles([jsonUpload.files[0], weightsUpload.files[0]]));

Example 4. Load a model from an HTTP server.

const model = await
     tf.loadModel('https://storage.googleapis.com/tfjs-models/tfjs/iris_v1/model.json')

Copy a model from one URL to another.

This function supports:

1. Copying within a storage medium, e.g., tf.io.copyModel('localstorage://model-1', 'localstorage://model-2')
2. Copying between two storage mediums, e.g., tf.io.copyModel('localstorage://model-1', 'indexeddb://model-1')

// First create and save a model.
const model = tf.sequential();
model.add(tf.layers.dense(
     {units: 1, inputShape: [10], activation: 'sigmoid'}));
await model.save('localstorage://demo/management/model1');

// Then list existing models.
console.log(JSON.stringify(await tf.io.listModels()));

// Copy the model, from Local Storage to IndexedDB.
await tf.io.copyModel(
     'localstorage://demo/management/model1',
     'indexeddb://demo/management/model1');

// List models again.
console.log(JSON.stringify(await tf.io.listModels()));

// Remove both models.
await tf.io.removeModel('localstorage://demo/management/model1');
await tf.io.removeModel('indexeddb://demo/management/model1');

List all models stored in registered storage mediums.

For a web browser environment, the registered mediums are Local Storage and IndexedDB.

// First create and save a model.
const model = tf.sequential();
model.add(tf.layers.dense(
     {units: 1, inputShape: [10], activation: 'sigmoid'}));
await model.save('localstorage://demo/management/model1');

// Then list existing models.
console.log(JSON.stringify(await tf.io.listModels()));

// Delete the model.
await tf.io.removeModel('localstorage://demo/management/model1');

// List models again.
console.log(JSON.stringify(await tf.io.listModels()));

Move a model from one URL to another.

This function supports:

1. Moving within a storage medium, e.g., tf.io.moveModel('localstorage://model-1', 'localstorage://model-2')
2. Moving between two storage mediums, e.g., tf.io.moveModel('localstorage://model-1', 'indexeddb://model-1')

// First create and save a model.
const model = tf.sequential();
model.add(tf.layers.dense(
     {units: 1, inputShape: [10], activation: 'sigmoid'}));
await model.save('localstorage://demo/management/model1');

// Then list existing models.
console.log(JSON.stringify(await tf.io.listModels()));

// Move the model, from Local Storage to IndexedDB.
await tf.io.moveModel(
     'localstorage://demo/management/model1',
     'indexeddb://demo/management/model1');

// List models again.
console.log(JSON.stringify(await tf.io.listModels()));

// Remove the moved model.
await tf.io.removeModel('indexeddb://demo/management/model1');

Remove a model specified by URL from a reigstered storage medium.

// First create and save a model.
const model = tf.sequential();
model.add(tf.layers.dense(
     {units: 1, inputShape: [10], activation: 'sigmoid'}));
await model.save('localstorage://demo/management/model1');

// Then list existing models.
console.log(JSON.stringify(await tf.io.listModels()));

// Delete the model.
await tf.io.removeModel('localstorage://demo/management/model1');

// List models again.
console.log(JSON.stringify(await tf.io.listModels()));

Exponetial Linear Unit (ELU).

It follows: f(x) = alpha * (exp(x) - 1.) for x < 0, f(x) = x for x >= 0.

Input shape: Arbitrary. Use the configuration inputShape when using this layer as the first layer in a model.

Output shape: Same shape as the input.

References:

• Fast and Accurate Deep Network Learning by Exponential Linear Units (ELUs)

Leaky version of a rectified linear unit.

It allows a small gradient when the unit is not active: f(x) = alpha * x for x < 0. f(x) = x for x >= 0.

Input shape: Arbitrary. Use the configuration inputShape when using this layer as the first layer in a model.

Output shape: Same shape as the input.

Softmax activation layer.

Input shape: Arbitrary. Use the configuration inputShape when using this layer as the first layer in a model.

Output shape: Same shape as the input.

Thresholded Rectified Linear Unit.

It follows: f(x) = x for x > theta, f(x) = 0 otherwise.

Input shape: Arbitrary. Use the configuration inputShape when using this layer as the first layer in a model.

Output shape: Same shape as the input.

References:

• Zero-Bias Autoencoders and the Benefits of Co-Adapting Features

Applies an activation function to an output.

This layer applies element-wise activation function. Other layers, notably dense can also apply activation functions. Use this isolated activation function to extract the values before and after the activation. For instance:

const input = tf.input({shape: [5]});
const denseLayer = tf.layers.dense({units: 1});
const activationLayer = tf.layers.activation({activation: 'relu6'});

// Obtain the output symbolic tensors by applying the layers in order.
const denseOutput = denseLayer.apply(input);
const activationOutput = activationLayer.apply(denseOutput);

// Create the model based on the inputs.
const model = tf.model({
     inputs: input,
     outputs: [denseOutput, activationOutput]
});

// Collect both outputs and print separately.
const [denseOut, activationOut] = model.predict(tf.randomNormal([6, 5]));
denseOut.print();
activationOut.print();

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.

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.

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].

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(JSON.stringify(flattenLayer.apply(input).shape));

Repeats the input n times in a new dimension.

  const model = tf.sequential();
  model.add(tf.layers.repeatVector({n: 4, inputShape: [2]}));
  const x = tf.tensor2d([[10, 20]]);
  // Use the model to do inference on a data point the model hasn't see
  model.predict(x).print();
  // output shape is now [batch, 2, 4]

Reshapes an input to a certain shape.

const input = tf.input({shape: [4, 3]});
const reshapeLayer = tf.layers.reshape({targetShape: [2, 6]});
// Inspect the inferred output shape of the Reshape layer, which
// equals `[null, 2, 6]`. (The 1st dimension is the undermined batch size.)
console.log(JSON.stringify(reshapeLayer.apply(input).shape));

Input shape: Arbitrary: although all dimensions in the input shape must be fixed. Use the ReshapeLayerConfig field input_shape when using this layer as the first layer in a model.

Output shape: [batchSize, targetShape[0], targetShape[1], ..., targetShape[targetShape.length - 1]].

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.

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'.

Transposed convolutional layer (sometimes called Deconvolution).

The need for transposed convolutions generally arises from the desire to use a transformation going in the opposite direction of a normal convolution, i.e., from something that has the shape of the output of some convolution to something that has the shape of its input while maintaining a connectivity pattern that is compatible with said convolution.

When using this layer as the first layer in a model, provide the configuration inputShape (Array of integers, does not include the sample axis), e.g., inputShape: [128, 128, 3] for 128x128 RGB pictures in dataFormat: 'channelsLast'.

Input shape: 4D tensor with shape: [batch, channels, rows, cols] if dataFormat is 'channelsFirst'. or 4D tensor with shape [batch, rows, cols, channels] if dataFormat is 'channelsLast.

Output shape: 4D tensor with shape: [batch, filters, newRows, newCols] if dataFormat is 'channelsFirst'. or 4D tensor with shape: [batch, newRows, newCols, filters] if dataFormat is 'channelsLast'.

References:

• A guide to convolution arithmetic for deep learning
• Deconvolutional Networks

Cropping layer for 2D input (e.g., image).

This layer can crop an input at the top, bottom, left and right side of an image tensor.

Input shape: 4D tensor with shape:

• If dataFormat is "channelsLast": [batch, rows, cols, channels]
• If data_format is "channels_first": [batch, channels, rows, cols].

Output shape: 4D with shape:

• If dataFormat is "channelsLast": [batch, croppedRows, croppedCols, channels]
• If dataFormat is "channelsFirst": [batch, channels, croppedRows, croppedCols].

Examples

const model = tf.sequential();
model.add(tf.layers.cropping2D({cropping:[[2, 2], [2, 2]],
                                inputShape: [128, 128, 3]}));
//now output shape is [batch, 124, 124, 3]

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.

Depthwise separable 2D convolution.

Separable convolution consists of first performing a depthwise spatial convolution (which acts on each input channel separately) followed by a pointwise convolution which mixes together the resulting output channels. The depthMultiplier argument controls how many output channels are generated per input channel in the depthwise step.

Intuitively, separable convolutions can be understood as a way to factorize a convolution kernel into two smaller kernels, or as an extreme version of an Inception block.

Input shape: 4D tensor with shape: [batch, channels, rows, cols] if data_format='channelsFirst' or 4D tensor with shape: [batch, rows, cols, channels] if data_format='channelsLast'.

Output shape: 4D tensor with shape: [batch, filters, newRows, newCols] if data_format='channelsFirst' or 4D tensor with shape: [batch, newRows, newCols, filters] if data_format='channelsLast'. rows and cols values might have changed due to padding.

Upsampling layer for 2D inputs.

Repeats the rows and columns of the data by size[0] and size[1] respectively.

Input shape: 4D tensor with shape: - If dataFormat is "channelsLast": [batch, rows, cols, channels] - If dataFormat is "channelsFirst": [batch, channels, rows, cols]

Output shape: 4D tensor with shape: - If dataFormat is "channelsLast": [batch, upsampledRows, upsampledCols, channels] - If dataFormat is "channelsFirst": [batch, channels, upsampledRows, upsampledCols]

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(JSON.stringify(sum.shape));
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.

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(JSON.stringify(average.shape));
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.

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(JSON.stringify(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).

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(JSON.stringify(max.shape));
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.

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(JSON.stringify(min.shape));
// You get [null, 2, 2], with the first dimension as the undetermined batch
// dimension.

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.

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:

• Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift

Average pooling operation for spatial data.

Input shape: [batchSize, inLength, channels]

Output shape: [batchSize, pooledLength, channels]

tf.avgPool1d is an alias.

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]

tf.avgPool2d is an alias.

Global average pooling operation for temporal data.

Input Shape: 3D tensor with shape: [batchSize, steps, features].

Output Shape:2D tensor with shape: [batchSize, features].

Global average pooling operation for spatial data.

Input shape:

• If dataFormat is CHANNEL_LAST: 4D tensor with shape: [batchSize, rows, cols, channels].
• If dataFormat is CHANNEL_FIRST: 4D tensor with shape: [batchSize, channels, rows, cols].

Output shape: 2D tensor with shape: [batchSize, channels].

Global max pooling operation for temporal data.

Input Shape: 3D tensor with shape: [batchSize, steps, features].

Output Shape:2D tensor with shape: [batchSize, features].

Global max pooling operation for spatial data.

Input shape:

• If dataFormat is CHANNEL_LAST: 4D tensor with shape: [batchSize, rows, cols, channels].
• If dataFormat is CHANNEL_FIRST: 4D tensor with shape: [batchSize, channels, rows, cols].

Output shape: 2D tensor with shape: [batchSize, channels].

Max pooling operation for temporal data.

Input shape: [batchSize, inLength, channels]

Output shape: [batchSize, pooledLength, channels]

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]

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(JSON.stringify(output.shape));
// [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.

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(JSON.stringify(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(JSON.stringify(output.shape));
// [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.

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 = lstm.apply(input);

console.log(JSON.stringify(output.shape));
// [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.

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(JSON.stringify(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(JSON.stringify(output.shape));
// [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.

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.

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(JSON.stringify(output.shape));
// [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.

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(JSON.stringify(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(JSON.stringify(output.shape));
// [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.

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.

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})}));
console.log(JSON.stringify(model.outputs[0].shape));
// 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],
}));
console.log(JSON.stringify(model.outputs[0].shape));

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();

Zero-padding layer for 2D input (e.g., image).

This layer can add rows and columns of zeros at the top, bottom, left and right side of an image tensor.

Input shape: 4D tensor with shape:

• If dataFormat is "channelsLast": [batch, rows, cols, channels]
• If data_format is "channels_first": [batch, channels, rows, cols].

Output shape: 4D with shape:

• If dataFormat is "channelsLast": [batch, paddedRows, paddedCols, channels]
• If dataFormat is "channelsFirst": [batch, channels, paddedRows, paddedCols].

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)

Divides two tf.Tensors element-wise, A / B. Supports broadcasting. The result is rounded with floor function.

const a = tf.tensor1d([1, 4, 9, 16]);
const b = tf.tensor1d([1, 2, 3, 4]);

a.floorDiv(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.floorDiv(b).print();  // or tf.floorDiv(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)

Returns the mod of a and b element-wise. floor(x / y) * y + mod(x, y) = x Supports broadcasting.

We also expose modStrict 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.mod(b).print();  // or tf.mod(a, b)

// Broadcast a mod b.
const a = tf.tensor1d([2, 4, 6, 8]);
const b = tf.scalar(5);

a.mod(b).print();  // or tf.mod(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. The result's dtype will be the upcasted type of the base and tf.exp() dtypes.

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).

Returns (a - b) * (a - b) element-wise. Supports broadcasting.

We also expose squaredDifferenceStrict 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.squaredDifference(b).print();  // or tf.squaredDifference(a, b)

// Broadcast squared difference  a with b.
const a = tf.tensor1d([2, 4, 6, 8]);
const b = tf.scalar(5);

a.squaredDifference(b).print();  // or tf.squaredDifference(a, b)

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)

Computes the inverse hyperbolic cos of the input tf.Tensor element-wise: acosh(x)

const x = tf.tensor1d([10, 1, 3, 5.7]);

x.acosh().print();  // or tf.acosh(x)

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)

Computes inverse hyperbolic sin of the input tf.Tensor element-wise: asinh(x)

const x = tf.tensor1d([0, 1, -1, .7]);

x.asinh().print();  // or tf.asinh(x)

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)