Package ai.djl.nn.norm

Class LayerNorm

  • All Implemented Interfaces:
    Block
    public class LayerNorm
extends AbstractBlock
    Layer normalization works by normalizing the values of input data for each input sample to have mean of 0 and variance of 1. Since this may alter the representation of a layer, two parameters (\ (\gamma\) and \(\beta\)) are learned along the normalization process to respectively scale and shift the normalized output (activations) to have any mean and variance so the network can utilize non-linear transformations such as sigmoid function as described in the paper. During backpropagation, both \(\gamma\) and \(\beta\) parameters are included following the chain-rule in derivation.

    Citing the abstract of the paper: "Training state-of-the-art, deep neural networks is computationally expensive. One way to reduce the training time is to normalize the activities of the neurons. A recently introduced technique called batch normalization uses the distribution of the summed input to a neuron over a mini-batch of training cases to compute a mean and variance which are then used to normalize the summed input to that neuron on each training case. This significantly reduces the training time in feed-forward neural networks. However, the effect of batch normalization is dependent on the mini-batch size and it is not obvious how to apply it to recurrent neural networks. In this paper, we transpose batch normalization into layer normalization by computing the mean and variance used for normalization from all of the summed inputs to the neurons in a layer on a single training case. Like batch normalization, we also give each neuron its own adaptive bias and gain which are applied after the normalization but before the non-linearity. Unlike batch normalization, layer normalization performs exactly the same computation at training and test times. It is also straightforward to apply to recurrent neural networks by computing the normalization statistics separately at each time step. Layer normalization is very effective at stabilizing the hidden state dynamics in recurrent networks. Empirically, we show that layer normalization can substantially reduce the training time compared with previously published techniques."

    • Field Detail

      • epsilon

        protected float epsilon

      • normalizedShape

        protected Shape normalizedShape

      • center

        protected boolean center

      • scale

        protected boolean scale

      • axis

        protected int[] axis

    • Method Detail

      • layerNorm

        public static NDList layerNorm​(NDArray input,
                               Shape normalizedShape,
                               NDArray gamma,
                               NDArray beta,
                               float eps)
        Applies Layer Normalization with average and variance for each input sample across the axis dimensions.
        Parameters:
        input - the input NDArray of shape (batchSize, inputChannel, *), * could be empty, width, (height, width), (depth, height, width)
        normalizedShape - dimensions to calculate average and variance from
        gamma - gamma weight NDArray
        beta - beta weight NDArray
        eps - a value added to the denominator for numerical stability
        Returns:
        the output NDArray of shape (batchSize, inputChannel, *), * could be empty, width, (height, width), (depth, height, width)

      • builder

        public static LayerNorm.Builder builder()
        Creates a builder to build a LayerNorm.
        Returns:
        a new builder

      • getOutputShapes

        public Shape[] getOutputShapes​(Shape[] inputShapes)
        Returns the expected output shapes of the block for the specified input shapes.
        Parameters:
        inputShapes - the shapes of the inputs
        Returns:
        the expected output shapes of the block

      • beforeInitialize

        protected void beforeInitialize​(Shape... inputShapes)
        Performs any action necessary before initialization. For example, keep the input information or verify the layout.
        Overrides:
        beforeInitialize in class AbstractBaseBlock
        Parameters:
        inputShapes - the expected shapes of the input

      • saveMetadata

        protected void saveMetadata​(java.io.DataOutputStream os)
                     throws java.io.IOException
        Override this method to save additional data apart from parameter values.

        This default implementation saves the currently set input shapes.

        Overrides:
        saveMetadata in class AbstractBaseBlock
        Parameters:
        os - the non-null output stream the parameter values and metadata are written to
        Throws:
        java.io.IOException - saving failed

      • loadMetadata

        public void loadMetadata​(byte loadVersion,
                         java.io.DataInputStream is)
                  throws java.io.IOException,
                         MalformedModelException
        Overwrite this to load additional metadata with the parameter values.

        If you overwrite AbstractBaseBlock.saveMetadata(DataOutputStream) or need to provide backward compatibility to older binary formats, you probably need to overwrite this. This default implementation checks if the version number fits, if not it throws an MalformedModelException. After that it restores the input shapes.

        Overrides:
        loadMetadata in class AbstractBaseBlock
        Parameters:
        loadVersion - the version used for loading this metadata.
        is - the input stream we are loading from
        Throws:
        java.io.IOException - loading failed
        MalformedModelException - data can be loaded but has wrong format