Uses of Package
org.deeplearning4j.nn.conf.layers

A neural network layer.

LSTM recurrent net, based on Graves: Supervised Sequence Labelling with Recurrent Neural Networks http://www.cs.toronto.edu/~graves/phd.pdf

Activation layer is a simple layer that applies the specified activation function to the input activations

Batch normalization layer
See: Ioffe and Szegedy, 2014, Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift https://arxiv.org/abs/1502.03167

Center loss is similar to triplet loss except that it enforces intraclass consistency and doesn't require feed forward of multiple examples.

2D Convolution layer (for example, spatial convolution over images).

Embedding layer: feed-forward layer that expects single integers per example as input (class numbers, in range 0 to numClass-1) as input.

Created by jeffreytang on 7/21/15.

Global pooling layer - used to do pooling over time for RNNs, and 2d pooling for CNNs.

Deprecated. 

Deprecated. 

LSTM recurrent neural network layer without peephole connections.

Output layer used for training via backpropagation based on labels and a specified loss function.

Pooling type:
MAX: Max pooling - output is the maximum value of the input values
AVG: Average pooling - output is the average value of the input values
SUM: Sum pooling - output is the sum of the input values
PNORM: P-norm pooling

A neural network layer.

A neural network layer.

The "PREFER_FASTEST" mode will pick the fastest algorithm for the specified parameters from the ConvolutionLayer.FwdAlgo, ConvolutionLayer.BwdFilterAlgo, and ConvolutionLayer.BwdDataAlgo lists, but they may be very memory intensive, so if weird errors occur when using cuDNN, please try the "NO_WORKSPACE" mode.

An optional dataFormat: "NDHWC" or "NCDHW".

LSTM recurrent net, based on Graves: Supervised Sequence Labelling with Recurrent Neural Networks http://www.cs.toronto.edu/~graves/phd.pdf

Activation layer is a simple layer that applies the specified activation function to the input activations

Autoencoder layer.

A neural network layer.

Upsampling base layer

Batch normalization layer
See: Ioffe and Szegedy, 2014, Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift https://arxiv.org/abs/1502.03167

Center loss is similar to triplet loss except that it enforces intraclass consistency and doesn't require feed forward of multiple examples.

3D Convolutional Neural Network Loss Layer.

Convolutional Neural Network Loss Layer.

1D (temporal) convolutional layer.

3D convolution layer configuration

An optional dataFormat: "NDHWC" or "NCDHW".

The "PREFER_FASTEST" mode will pick the fastest algorithm for the specified parameters from the ConvolutionLayer.FwdAlgo, ConvolutionLayer.BwdFilterAlgo, and ConvolutionLayer.BwdDataAlgo lists, but they may be very memory intensive, so if weird errors occur when using cuDNN, please try the "NO_WORKSPACE" mode.

The backward filter algorithm to use when ConvolutionLayer.AlgoMode is set to "USER_SPECIFIED".

2D deconvolution layer configuration
Deconvolutions are also known as transpose convolutions or fractionally strided convolutions.

Dense layer: a standard fully connected feed forward layer

2D depth-wise convolution layer configuration.

Dropout layer.

Deprecated. 

A neural network layer.

Implements Dot Product Self Attention with learned queries Takes in RNN style input in the shape of [batchSize, features, timesteps] and applies dot product attention using learned queries.

SameDiff version of a 1D locally connected layer.

SameDiff version of a 2D locally connected layer.

Local response normalization layer
See section 3.3 of http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf

LossLayer is a flexible output layer that performs a loss function on an input without MLP logic.

LSTM recurrent neural network layer without peephole connections.

Output layer used for training via backpropagation based on labels and a specified loss function.

Pooling type:
MAX: Max pooling - output is the maximum value of the input values
AVG: Average pooling - output is the average value of the input values
SUM: Sum pooling - output is the sum of the input values
PNORM: P-norm pooling

Implements Recurrent Dot Product Attention Takes in RNN style input in the shape of [batchSize, features, timesteps] and applies dot product attention using the hidden state as the query and all time steps as keys/values.

Recurrent Neural Network Loss Layer.

A version of OutputLayer for recurrent neural networks.

1D (temporal) subsampling layer - also known as pooling layer.

3D subsampling / pooling layer for convolutional neural networks Supports max and average pooling modes

Subsampling layer also referred to as pooling in convolution neural nets Supports the following pooling types: MAX, AVG, SUM, PNORM

Upsampling 1D layer
Repeats each step size times along the temporal/sequence axis (dimension 2)
For input shape [minibatch, channels, sequenceLength] output has shape [minibatch, channels, size * sequenceLength]
Example: If input (for a single example, with channels down page, and sequence from left to right) is: [ A1, A2, A3] [ B1, B2, B3] Then output with size = 2 is: [ A1, A1, A2, A2, A3, A3] [ B1, B1, B2, B2, B3, B2]

Upsampling 2D layer
Repeats each value (or rather, set of depth values) in the height and width dimensions by size[0] and size[1] times respectively.

Upsampling 3D layer
Repeats each value (all channel values for each x/y/z location) by size[0], size[1] and size[2]
If input has shape [minibatch, channels, depth, height, width] then output has shape [minibatch, channels, size[0] * depth, size[1] * height, size[2] * width]

Zero padding 1D layer for convolutional neural networks.

Zero padding layer for convolutional neural networks (2D CNNs).

A neural network layer.

A neural network layer.

Created by jeffreytang on 7/21/15.

A neural network layer.

A neural network layer.

A neural network layer.

Created by jeffreytang on 7/21/15.

A neural network layer.

A neural network layer.

A neural network layer.

A neural network layer.

Created by jeffreytang on 7/21/15.

A neural network layer.

A neural network layer.

A neural network layer.

Created by jeffreytang on 7/21/15.

A neural network layer.

An optional dataFormat: "NDHWC" or "NCDHW".

A neural network layer.

A neural network layer.

A neural network layer.

The "PREFER_FASTEST" mode will pick the fastest algorithm for the specified parameters from the ConvolutionLayer.FwdAlgo, ConvolutionLayer.BwdFilterAlgo, and ConvolutionLayer.BwdDataAlgo lists, but they may be very memory intensive, so if weird errors occur when using cuDNN, please try the "NO_WORKSPACE" mode.

The backward filter algorithm to use when ConvolutionLayer.AlgoMode is set to "USER_SPECIFIED".

Pooling type:
MAX: Max pooling - output is the maximum value of the input values
AVG: Average pooling - output is the average value of the input values
SUM: Sum pooling - output is the sum of the input values
PNORM: P-norm pooling

The "PREFER_FASTEST" mode will pick the fastest algorithm for the specified parameters from the ConvolutionLayer.FwdAlgo, ConvolutionLayer.BwdFilterAlgo, and ConvolutionLayer.BwdDataAlgo lists, but they may be very memory intensive, so if weird errors occur when using cuDNN, please try the "NO_WORKSPACE" mode.

The backward filter algorithm to use when ConvolutionLayer.AlgoMode is set to "USER_SPECIFIED".

Pooling type:
MAX: Max pooling - output is the maximum value of the input values
AVG: Average pooling - output is the average value of the input values
SUM: Sum pooling - output is the sum of the input values
PNORM: P-norm pooling

A neural network layer.

LSTM recurrent net, based on Graves: Supervised Sequence Labelling with Recurrent Neural Networks http://www.cs.toronto.edu/~graves/phd.pdf

Deprecated. 

A neural network layer.

A neural network layer.

Created by jeffreytang on 7/21/15.

Activation layer is a simple layer that applies the specified activation function to the input activations

1D (temporal) convolutional layer.

2D deconvolution layer configuration
Deconvolutions are also known as transpose convolutions or fractionally strided convolutions.

2D Separable convolution layer configuration.

Upsampling 1D layer
Repeats each step size times along the temporal/sequence axis (dimension 2)
For input shape [minibatch, channels, sequenceLength] output has shape [minibatch, channels, size * sequenceLength]
Example: If input (for a single example, with channels down page, and sequence from left to right) is: [ A1, A2, A3] [ B1, B2, B3] Then output with size = 2 is: [ A1, A1, A2, A2, A3, A3] [ B1, B1, B2, B2, B3, B2]

Upsampling 3D layer
Repeats each value (all channel values for each x/y/z location) by size[0], size[1] and size[2]
If input has shape [minibatch, channels, depth, height, width] then output has shape [minibatch, channels, size[0] * depth, size[1] * height, size[2] * width]

Zero padding 3D layer for convolutional neural networks.

Activation layer is a simple layer that applies the specified activation function to the input activations

Dropout layer.

Local response normalization layer
See section 3.3 of http://www.cs.toronto.edu/~fritz/absps/imagenet.pdf

Embedding layer for sequences: feed-forward layer that expects fixed-length number (inputLength) of integers/indices per example as input, ranged from 0 to numClasses - 1.

SameDiff version of a 1D locally connected layer.

Dropout layer.

Batch normalization layer
See: Ioffe and Szegedy, 2014, Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift https://arxiv.org/abs/1502.03167

Global pooling layer - used to do pooling over time for RNNs, and 2d pooling for CNNs.

1D (temporal) subsampling layer - also known as pooling layer.

Subsampling layer also referred to as pooling in convolution neural nets Supports the following pooling types: MAX, AVG, SUM, PNORM

A neural network layer.

A neural network layer.

A neural network layer.

The "PREFER_FASTEST" mode will pick the fastest algorithm for the specified parameters from the ConvolutionLayer.FwdAlgo, ConvolutionLayer.BwdFilterAlgo, and ConvolutionLayer.BwdDataAlgo lists, but they may be very memory intensive, so if weird errors occur when using cuDNN, please try the "NO_WORKSPACE" mode.

An optional dataFormat: "NDHWC" or "NCDHW".

Pooling type:
MAX: Max pooling - output is the maximum value of the input values
AVG: Average pooling - output is the average value of the input values
SUM: Sum pooling - output is the sum of the input values
PNORM: P-norm pooling

Batch normalization layer
See: Ioffe and Szegedy, 2014, Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift https://arxiv.org/abs/1502.03167

Dense layer: a standard fully connected feed forward layer