batchnorm1d_layer_type

batchnorm1d_layer_type(
   input_shape,
   num_channels,
   num_inputs,
   batch_size,
   momentum=0.99,
   epsilon=0.001,
   gamma_init_mean=1.0,
   gamma_init_std=0.01,
   beta_init_mean=0.0,
   beta_init_std=0.01,
   kernel_initialiser='ones',
   bias_initialiser='zeros',
   moving_mean_initialiser='zeros',
   moving_variance_initialiser='ones'
)

THIS LAYER TYPE WILL ONLY BE AVAILABLE IN VERSION 1.3.0 AND LATER

The batchnorm1d_layer_type derived type provides 0D and 1D batch normalisation layer. If input_shape is rank 1 (or only num_channels or num_inputs is given), then the input and output data shape will be CxB, where C = number of channels/inputs, and B = batch size. If input_shape is rank 2 (or both num_channels and num_inputs is given), then the input and output data shape will be IxCxB, where I = number of features per channel, C = number of channels, and B = batch size.

This layer normalises over samples within a batch to improve stability of data.

Arguments

• input_shape: The shape of the input data for one sample. This is required only if this layer is the first (non-input) layer of the network.
• num_inputs: An integer scalar. Defines the number of input features per channel.
• num_channels: An integer scalar. Defines the number of channels in the input data.
• batch_size: Integer. The number samples in a batch. This is optional (the enclosing network structure can handle it instead).
• momentum: A real scalar. Defines the speed at which learning is performed.
• epsilon: A real scalar. A small positive number that is used to avoid division by zero.
• gamma_init_mean: The mean of initial gamma values (for gaussian random initialisation).
• gamma_init_std: The standard deviation of initial gamma values (for gaussian random initialisation).
• beta_init_mean: The mean of initial beta values (for gaussian random initialisation).
• beta_init_std: The standard deviation of initial beta values (for gaussian random initialisation).
• kernel_initialiser: Initialiser for the kernel weights, i.e. gamma (see Initialisers).
• bias_initialiser: Initialiser for the biases, i.e. beta (see Initialisers).
• moving_mean_initialiser: Initialiser for the moving mean, for momentum-based learning (see Initialisers).
• moving_variance_initialiser: Initialiser for the moving variance, for momentum-based learning (see Initialisers).