Project an input of shape (T, dim) to (T, num_heads, head_dim) using a kernel of shape (dim, num_heads, head_dim) and an optional bias of shape (num_h

Serialize AdditiveAttention layer to dict

Serialize Attention layer to dict

Serialize batch normalization layer to dict

Serialize Bidirectional wrapper around an LSTM/GRU/SimpleRNN to dict

Serialize CategoryEncoding layer to dict

Serialize Conv1D layer to dict

Serialize Conv1D transpose layer to dict

Serialize Conv2D layer to dict

Serialize Conv2D transpose layer to dict

Serialize Conv3D layer to dict

Serialize Conv3DTranspose layer to dict

Serialize ConvLSTM1D/2D layer to dict

Serialize dense layer to dict

Serialize DepthwiseConv1D layer to dict by promoting to a 2D depthwise representation.

Serialize DepthwiseConv2D layer to dict

Check valid configuration of Dot layer

Serialize EinsumDense layer to dict

Serialize Embedding layer to dict

Serialize GroupNormalization layer to dict

Serialize GroupedQueryAttention layer to dict

Serialize GRU layer to dict

Serialize input layer to dict

Serialize layer normalization layer to dict

Serialize LSTM layer to dict

Serialize MultiHeadAttention layer to dict

Serialize normalization layer to dict

Serialize prelu layer to dict

Serialize Rescaling layer to dict

Serialize Resizing layer to dict

Serialize RMSNormalization layer to dict

Serialize a generic RNN layer wrapping LSTMCell/GRUCell/SimpleRNNCell or StackedRNNCells.

Serialize SeparableConv2D layer to dict

Serialize SimpleRNN layer to dict

Serialize softmax layer to dict

Serialize 3-tensor to a dict

Serialize UpSampling2D layer to dict

Validate UpSampling3D layer config (size is taken from layer config)

Converts a memory block holding 8-bit values into a tensor. Data must be stored row-wise (and channels_last). Scales the values from range [0, 255] in

The dimensions are right-aligned (left-padded) compared to Keras. I.e., if you have a position (or shape) of (a, b) in Keras it corresponds to (0, 0,

The outer (left-most) dimensions are not used for batch prediction. If you like to do multiple forward passes on a model at once, use fdeep::model::pr

Converts a tensor into a vector of bytes. Data will be stored row-wise (and channels_last). Scales the values from range [low, high] into [0, 255].